Predicting Diagnostic Progression to Schizophrenia or Bipolar Disorder via Machine Learning
Open accessAbstract
Importance: The diagnosis of schizophrenia and bipolar disorder is often delayed several years despite illness typically emerging in late adolescence or early adulthood, which impedes initiation of targeted treatment. Objective: To investigate whether machine learning models trained on routine clinical data from electronic health records (EHRs) can predict diagnostic progression to schizophrenia or bipolar disorder among patients undergoing treatment in psychiatric services for other mental illness. Design, Setting, and Participants: This cohort study was based on data from EHRs from the Psychiatric Services of the Central Denmark Region. All patients aged 15 to 60 years with at least 2 contacts (at least 3 months apart) with the Psychiatric Services of the Central Denmark Region between January 1, 2013, and November 21, 2016, were included. Analysis occurred from December 2022 to November 2024. Exposures: Predictors based on EHR data, including medications, diagnoses, and clinical notes. Main Outcomes and Measures: Diagnostic transition to schizophrenia or bipolar disorder within 5 years, predicted 1 day before outpatient contacts by means of elastic net regularized logistic regression and extreme gradient boosting (XGBoost) models. The area under the receiver operating characteristic curve (AUROC) was used to determine the best performing model. Results: The study included 24 449 patients (median [Q1-Q3] age at time of prediction, 32.2 [24.2-42.5] years; 13 843 female [56.6%]) and 398 922 outpatient contacts. Transition to the first occurrence of either schizophrenia or bipolar disorder was predicted by the XGBoost model, with an AUROC of 0.70 (95% CI, 0.70-0.70) on the training set and 0.64 (95% CI, 0.63-0.65) on the test set, which consisted of 2 held-out hospital sites. At a predicted positive rate of 4%, the XGBoost model had a sensitivity of 9.3%, a specificity of 96.3%, and a positive predictive value (PPV) of 13.0%. Predicting schizophrenia separately yielded better performance (AUROC, 0.80; 95% CI, 0.79-0.81; sensitivity, 19.4%; specificity, 96.3%; PPV, 10.8%) than was the case for bipolar disorder (AUROC, 0.62, 95% CI, 0.61-0.63; sensitivity, 9.9%; specificity, 96.2%; PPV, 8.4%). Clinical notes proved particularly informative for prediction. Conclusions and Relevance: These findings suggest that it is possible to predict diagnostic transition to schizophrenia and bipolar disorder from routine clinical data extracted from EHRs, with schizophrenia being notably easier to predict than bipolar disorder.
Verified 2026-10-09 · 2 references not checked yet
AI snapshot
Written from the paper’s full text.
TL;DR
This cohort study used routine electronic health record data from the Psychiatric Services of the Central Denmark Region to predict 5-year diagnostic progression to schizophrenia or bipolar disorder among psychiatric outpatients. XGBoost models predicted schizophrenia substantially better than bipolar disorder, with clinical-note text features proving particularly informative.
Key findings
- Among 24,449 patients and 398,922 outpatient contacts, the joint XGBoost model predicting first occurrence of schizophrenia or bipolar disorder achieved an AUROC of 0.64 (95% CI, 0.63-0.65) on the held-out test set; at a 4% predicted positive rate, sensitivity was 9.3%, specificity 96.3%, and PPV 13.0%.
- Predicting schizophrenia separately performed better: AUROC 0.80 (95% CI, 0.79-0.81); at a 4% predicted positive rate, sensitivity 19.4%, specificity 96.3%, and PPV 10.8%.
- Bipolar disorder was more difficult to predict: AUROC 0.62 (95% CI, 0.61-0.63); at a 4% predicted positive rate, sensitivity 9.9%, specificity 96.2%, and PPV 8.4%.
- Clinical notes were particularly informative; text-only models performed practically equivalent to models using structured plus text features, and top features included Danish tokens related to discharge, admission, voices, and inpatient activities.
Methods
- Cohort study using EHRs from the Psychiatric Services of the Central Denmark Region, including patients aged 15 to 60 years with at least 2 contacts at least 3 months apart between January 1, 2013, and November 21, 2016; analysis occurred from December 2022 to November 2024.
- Predictions were issued 1 day before scheduled outpatient contacts with a 5-year look-ahead window; outcomes were first ICD-10 diagnosis of schizophrenia (F20) or schizoaffective disorder (F25), or manic episode (F30) or bipolar affective disorder (F31).
- Elastic net regularized logistic regression and XGBoost models were trained and hyperparameter tuned with Optuna; predictors included structured EHR data and clinical-note embeddings (TF-IDF with 1000 features performed best), with TabDDPM data augmentation; AUROC was the main evaluation metric, using a train/test split by hospital site.
Limitations
- The paper states that data were restricted to patients receiving psychiatric treatment and did not include primary care, so models are primarily useful for patients progressing from less severe mental disorders and may not generalize to broader, low-risk populations.
- The paper states that text models might fit to already present clinical suspicion and provide less lead time; test performance was evaluated on only 2 hospital sites; computational limitations precluded internal-external cross-validation; generalizability outside the Central Denmark Region is unknown; medication data covered only inpatient-administered medication, not prescribed medication; and temporal stability should be investigated before implementation.
- Evident from the reported results and discussion: there was a substantial drop from training to test performance (e.g., joint model AUROC 0.70 training vs 0.64 test; bipolar disorder model 0.67 training vs 0.62 test), suggesting distribution shifts across hospital sites, and the bipolar disorder model was less robust across calendar time, with a noticeable dip around the third quarter of 2015.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
0 of 24 cited sources checked · 0 supported · 0 flagged
- Findingown data§ Results
The text-feature set that provided the best predictive performance on the training set was TF-IDF with 1000 features, trained on all note types.
“The text-feature set that provided the best predictive performance on the training set was TF-IDF with 1000 features, trained on all note types (eTable 3 in Supplement 1).”
Quote in the paperSubstantial claim - Findingown data§ Results
The joint model achieved an AUROC of 0.57 (95% CI, 0.56-0.58) for the bipolar disorder-only outcome on the test set.
“The joint model achieved an AUROC of 0.74 (95% CI, 0.73-0.75) for the schizophrenia-only outcome and 0.57 (95% CI, 0.56-0.58) for the bipolar disorder–only outcome on the test set (eTable 9 in Supplement 1).”
Quote in the paperNot a substantial claim - Findingown data§ Results
The top 10 most important predictors by information gain in the best-performing text-only joint model were Udskrivelse, Spillet, Veninder, Stemmerne, Udskrives, Gav, Indlæggelsen, Spille, Morgenstunden, and Forklare, with information gains ranging from 0.003781 to 0.002103.
“Top 10 Most Important Predictors by Information Gain in the Best-Performing (Text-Only) Joint Model.”
Quote in the paperNot a substantial claim - Findingown data§ Results
Schizophrenia is notably easier to predict than bipolar disorder.
“These findings suggest that it is possible to predict diagnostic transition to schizophrenia and bipolar disorder from routine clinical data extracted from EHRs, with schizophrenia being notably easier to predict than bipolar disorder.”
Quote in the paperSubstantial claim - Findingown data§ Results
Text embeddings of words, including voices and admission, were found to be highly influential for the joint model.
“Notably, text embeddings of words, including voices and admission, were found to be highly influential for the model.”
Quote in the paperNot a substantial claim
- Findingown data§ Results
The bipolar disorder-only model achieved an AUROC of 0.62 (95% CI, 0.61-0.63) on the test set.
“Predicting schizophrenia separately yielded better performance (AUROC, 0.80; 95% CI, 0.79-0.81; sensitivity, 19.4%; specificity, 96.3%; PPV, 10.8%) than was the case for bipolar disorder (AUROC, 0.62, 95% CI, 0.61-0.63; sensitivity, 9.9%; specificity, 96.2%; PPV, 8.4%).”
Quote in the paperSubstantial claim - Findingown data§ Results
At a predicted positive rate of 4%, the XGBoost joint model had a specificity of 96.3%.
“The sensitivity at the level of prediction times was 9.3%, the specificity was 96.3%, and 13.5% of all patients who received a diagnosis of schizophrenia or bipolar disorder (47 of 347 patients) were predicted positive at least once (eTable 8 in Supplement 1).”
Quote in the paperNot a substantial claim - Findingown data§ Results
The joint model trained on the 85% split, stratified only by patient, obtained an AUROC of 0.72 (95% CI, 0.71-0.73) on the 15% test set using the text-only feature set and the optimal hyperparameters identified previously.
“The model (joint outcomes) trained on the 85% split, stratified only by patient, obtained an AUROC of 0.72 (95% CI, 0.71-0.73) on the 15% test set using the text-only feature set and the optimal hyperparameters identified previously.”
Quote in the paperSubstantial claim - Findingown data§ Results
The median (Q1-Q3) time from the first positive prediction to the outcome was 1.1 (0.2-2.2) years on the test set.
“The median (Q1-Q3) time from the first positive prediction to the outcome was 1.1 (0.2-2.2) years (see Figure 2D).”
Quote in the paperNot a substantial claim - Findingown data§ Discussion
Inspection of the most important words by feature importance and the context they appeared in within the EHRs revealed that many were related to hospital admission or psychiatric symptoms.
“Inspection of the most important words by feature importance and the context they appeared in within the EHRs revealed that many were related to hospital admission or psychiatric symptoms.”
Quote in the paperNot a substantial claim - Findingown data§ Results
Performance of the joint model was quite stable across levels of time from the first contact, with some instability at the extremes, likely owing partially to lack of data.
“Performance was quite stable across levels of time from the first contact, with some instability at the extremes, likely owing partially to lack of data.”
Quote in the paperNot a substantial claim - Findingown data§ Results
For the separate outcomes, inclusion of text and synthetic features seemed to improve performance on both the training and test set.
“As shown in eTable 12 in Supplement 1, for the separate outcomes, inclusion of text and synthetic features seemed to improve performance on both the training and test set.”
Quote in the paperNot a substantial claim - Findingown data§ Results
The data augmentation method that provided the best predictive performance on the training set was TabDDPM with a 2× multiplier for the minority class.
“The data augmentation method that provided the best predictive performance on the training set was TabDDPM with a 2 × multiplier for the minority class (eTable 4 in Supplement 1).”
Quote in the paperSubstantial claim - Findingown data§ Discussion
A substantial drop in performance was observed when moving from the training set to the test set, particularly for the joint model and the bipolar disorder model.
“A substantial drop in performance was observed when moving from the training set to the test set, particularly for the joint model and the bipolar disorder model.”
Quote in the paperSubstantial claim - Findingown data§ Results
The best model predicting schizophrenia achieved an AUROC of 0.80 (95% CI, 0.79-0.81) on the test set.
“The models predicting schizophrenia obtained the best performance, with an AUROC on the test set of 0.80 (95% CI, 0.79-0.81) for the best model.”
Quote in the paperSubstantial claim - Findingown data§ Results
For schizophrenia, the joint model achieved an AUROC of 0.75 (95% CI, 0.74-0.75) compared with 0.78 (95% CI, 0.77-0.78) for the schizophrenia-only model on the training set.
“Specifically, for schizophrenia, the joint model achieved an AUROC of 0.75 (95% CI, 0.74-0.75) compared with 0.78 (95% CI, 0.77-0.78) for the schizophrenia-only model.”
Quote in the paperSubstantial claim - Findingown data§ Results
At a predicted positive rate of 4%, the XGBoost joint model had a positive predictive value (PPV) of 13.0%.
“The PPV was 13.0%, indicating that for every 7.7 positive predictions, 1 prediction time was followed by a diagnosis of schizophrenia or bipolar disorder within 5 years.”
Quote in the paperSubstantial claim - Findingown data§ Results
At a threshold of the 4% highest risk predictions marked as positive, the median lead time for the best model to flag patients who would develop schizophrenia or bipolar disorder was 0.7 years on the training set.
“At a threshold of the 4% highest risk predictions marked as positive, the median lead time for the best model to flag patients who would develop schizophrenia or bipolar disorder was 0.7 years (on the training set).”
Quote in the paperNot a substantial claim - Findingown data§ Results
The best model predicting bipolar disorder achieved an AUROC of 0.62 (95% CI, 0.61-0.63) on the test set.
“Bipolar disorder proved more difficult, with the best model achieving an AUROC of 0.62 (95% CI, 0.61-0.63) on the test set.”
Quote in the paperSubstantial claim - Findingown data§ Results
At a predicted positive rate of 4%, the XGBoost joint model had a sensitivity of 9.3%.
“The sensitivity at the level of prediction times was 9.3%, the specificity was 96.3%, and 13.5% of all patients who received a diagnosis of schizophrenia or bipolar disorder (47 of 347 patients) were predicted positive at least once (eTable 8 in Supplement 1).”
Quote in the paperNot a substantial claim - Findingown data§ Results
The model for bipolar disorder was less robust, particularly across calendar time, with a noticeable dip in performance around the third quarter of 2015.
“As shown in eFigure 6 in Supplement 1, the model for bipolar disorder was less robust, particularly across calendar time, with a noticeable dip in performance around the third quarter of 2015.”
Quote in the paperNot a substantial claim - Findingown data§ Results
The joint model's performance approached the performance of the separate models when evaluated on each outcome separately on the training set.
“The performance of the joint model (ie, the model trained to predict the first occurring onset of either schizophrenia or bipolar disorder) approached the performance of the separate models when evaluated on each outcome separately on the training set.”
Quote in the paperNot a substantial claim - Findingown data§ Results
The schizophrenia-only model achieved an AUROC of 0.80 (95% CI, 0.79-0.81) on the test set.
“Predicting schizophrenia separately yielded better performance (AUROC, 0.80; 95% CI, 0.79-0.81; sensitivity, 19.4%; specificity, 96.3%; PPV, 10.8%) than was the case for bipolar disorder (AUROC, 0.62, 95% CI, 0.61-0.63; sensitivity, 9.9%; specificity, 96.2%; PPV, 8.4%).”
Quote in the paperSubstantial claim - Findingnone given§ Results
Clinical notes proved particularly informative for prediction.
“Clinical notes proved particularly informative for prediction.”
Quote in the paperNot a substantial claim - Findingown data§ Results
The feature set including structured data and text and the feature set only using text features both achieved an AUROC of 0.69 (95% CI, 0.69-0.69) on the training set.
“The feature set including structured data and text and the feature set only using text features both achieved an AUROC 0.69 (95% CI, 0.69-0.69).”
Quote in the paperNot a substantial claim - Findingown data§ Results
Predicting schizophrenia separately yielded better performance than predicting bipolar disorder separately.
“Predicting schizophrenia separately yielded better performance (AUROC, 0.80; 95% CI, 0.79-0.81; sensitivity, 19.4%; specificity, 96.3%; PPV, 10.8%) than was the case for bipolar disorder (AUROC, 0.62, 95% CI, 0.61-0.63; sensitivity, 9.9%; specificity, 96.2%; PPV, 8.4%).”
Quote in the paperSubstantial claim - Findingown data§ Results
13.5% of all patients who received a diagnosis of schizophrenia or bipolar disorder (47 of 347 patients) were predicted positive at least once.
“The sensitivity at the level of prediction times was 9.3%, the specificity was 96.3%, and 13.5% of all patients who received a diagnosis of schizophrenia or bipolar disorder (47 of 347 patients) were predicted positive at least once (eTable 8 in Supplement 1).”
Quote in the paperNot a substantial claim - Findingown data§ Results
For the joint model on the test set, sensitivity was generally higher for predicting schizophrenia compared with bipolar disorder.
“As shown in Figure 2C, for the joint model, sensitivity was generally higher for predicting schizophrenia compared with bipolar disorder.”
Quote in the paperNot a substantial claim - Findingown data§ Results
When applied to the test set, the XGBoost joint model using only text features that performed best achieved an AUROC of 0.64 (95% CI, 0.63-0.65).
“When applied to the test set (Figure 2), the XGBoost joint model using only text features that performed best achieved an AUROC of 0.64 (95% CI, 0.63-0.65) (Figure 2A).”
Quote in the paperSubstantial claim - Findingown data§ Results
The performance of the joint model was stable across sex and age.
“Figure 3 shows that the performance of the joint model was stable across sex and age.”
Quote in the paperNot a substantial claim - Findingown data§ Results
Using only structured data yielded an AUROC of 0.66 (95% CI, 0.65-0.66) on the training set.
“Using only structured data yielded an AUROC of 0.66 (95% CI, 0.65-0.66).”
Quote in the paperNot a substantial claim - Findingown data§ Results
XGBoost was superior to logistic regression in all cases.
“XGBoost was superior to logistic regression in all cases (eTable 6 in Supplement 1).”
Quote in the paperSubstantial claim - Findingcitation§ Discussion
Models trained with both structured and text-based predictors performed practically equivalent to models trained with only text-based features.
“Indeed, models trained with both structured and text-based predictors performed practically equivalent to models trained with only text-based features; this underscores the importance of text in clinical prediction modeling within psychiatry.12,31”
Quote in the paperSubstantial claim - Findingown data§ Results
The feature set including structured data, text, and synthetic data performed slightly better than the other feature sets on the training set, with an AUROC of 0.70 (95% CI, 0.70-0.70).
“Specifically, eTable 5 in Supplement 1 shows that the feature set including structured data, text, and synthetic data performed slightly better than the other feature sets on the training set, with an AUROC of 0.70 (95% CI, 0.70-0.70).”
Quote in the paperNot a substantial claim - Findingown data§ Results
The joint model achieved an AUROC of 0.74 (95% CI, 0.73-0.75) for the schizophrenia-only outcome on the test set.
“The joint model achieved an AUROC of 0.74 (95% CI, 0.73-0.75) for the schizophrenia-only outcome and 0.57 (95% CI, 0.56-0.58) for the bipolar disorder–only outcome on the test set (eTable 9 in Supplement 1).”
Quote in the paperSubstantial claim - Findingown data§ Results
For the best schizophrenia model on the test set, at a 4% predicted positive rate, sensitivity was 19.4%, specificity 96.3%, and PPV 10.8%.
“At a 4% predicted positive rate, sensitivity was 19.4%, specificity was 96.3%, and the PPV was 10.8% on the test set.”
Quote in the paperSubstantial claim - Findingown data§ Results
For the best bipolar disorder model on the test set, at a 4% predicted positive rate, sensitivity was 9.9%, specificity 96.2%, and PPV 8.4%.
“At a 4% predicted positive rate, sensitivity was 9.9%, specificity was 96.2%, and the PPV was 8.4% on the test set.”
Quote in the paperNot a substantial claim - Findingown data§ Results
For bipolar disorder, the joint model achieved an AUROC of 0.66 (95% CI, 0.65-0.66) compared with 0.67 (95% CI, 0.67-0.68) for the bipolar disorder-only model on the training set.
“For bipolar disorder, the joint model achieved an AUROC of 0.66 (95% CI, 0.65-0.66) compared with 0.67 (95% CI, 0.67-0.68) for the bipolar disorder–only model.”
Quote in the paperSubstantial claim - Premisecitation§ Introduction
Timely and accurate diagnosis is crucial because diagnostic delay impedes the initiation of targeted treatment.
“Timely and accurate diagnosis is crucial because diagnostic delay impedes the initiation of targeted treatment.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [4] Altamura AC, Buoli M, Caldiroli A, et al. Misdiagnosis, duration of untreated illness (DUI) and outcome in bipolar patients with psychotic symptoms: a naturalistic study. J Affect Disord. 2015;182:70-75. doi: 10.1016/j.jad.2015.04.024 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
- cites [5] Penttilä M, Jääskeläinen E, Hirvonen N, Isohanni M, Miettunen J. Duration of untreated psychosis as predictor of long-term outcome in schizophrenia: systematic review and meta-analysis. Br J Psychiatry. 2014;205(2):88-94. doi: 10.1192/bjp.bp.113.127753 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
- Premisecitation§ Introduction
The longer the duration of untreated illness, the worse the prognosis becomes.
“Furthermore, the longer the duration of untreated illness, the worse the prognosis becomes.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [4] Altamura AC, Buoli M, Caldiroli A, et al. Misdiagnosis, duration of untreated illness (DUI) and outcome in bipolar patients with psychotic symptoms: a naturalistic study. J Affect Disord. 2015;182:70-75. doi: 10.1016/j.jad.2015.04.024 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
- cites [5] Penttilä M, Jääskeläinen E, Hirvonen N, Isohanni M, Miettunen J. Duration of untreated psychosis as predictor of long-term outcome in schizophrenia: systematic review and meta-analysis. Br J Psychiatry. 2014;205(2):88-94. doi: 10.1192/bjp.bp.113.127753 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
- Premisecitation§ Introduction
Diagnosis of schizophrenia and bipolar disorder is often delayed several years despite typical emergence in late adolescence or early adulthood.
“Despite typically emerging in late adolescence or early adulthood, diagnosis is often delayed several years.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [4] Altamura AC, Buoli M, Caldiroli A, et al. Misdiagnosis, duration of untreated illness (DUI) and outcome in bipolar patients with psychotic symptoms: a naturalistic study. J Affect Disord. 2015;182:70-75. doi: 10.1016/j.jad.2015.04.024 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
- cites [5] Penttilä M, Jääskeläinen E, Hirvonen N, Isohanni M, Miettunen J. Duration of untreated psychosis as predictor of long-term outcome in schizophrenia: systematic review and meta-analysis. Br J Psychiatry. 2014;205(2):88-94. doi: 10.1192/bjp.bp.113.127753 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
- Interpretationown data§ Discussion
Text is important in clinical prediction modeling within psychiatry.
“Indeed, models trained with both structured and text-based predictors performed practically equivalent to models trained with only text-based features; this underscores the importance of text in clinical prediction modeling within psychiatry.12,31”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [12] Irving J, Patel R, Oliver D, et al. Using natural language processing on electronic health records to enhance detection and prediction of psychosis risk. Schizophr Bull. 2021;47(2):405-414. doi: 10.1093/schbul/sbaa126 [DOI] [PMC free article] [PubMed] [Google Scholar] source
Not checked in the free run.
- cites [31] Rumshisky A, Ghassemi M, Naumann T, et al. Predicting early psychiatric readmission with natural language processing of narrative discharge summaries. Transl Psychiatry. 2016;6(10):e921-e921. doi: 10.1038/tp.2015.182 [DOI] [PMC free article] [PubMed] [Google Scholar]
Not checked in the free run.
- Study design and conductreasoning§ Methods
The cohort was limited to outpatient contacts occurring after January 1, 2013, due to inconsistencies in the data before 2013.
“The cohort was limited to outpatient contacts occurring after January 1, 2013, due to inconsistencies in the data before 2013, stemming from the gradual implementation of a new EHR system in 2011.”
Quote in the paperSubstantial claim - Study design and conductnone given§ Discussion
The training and test sets contained data from different hospitals in the CDR.
“The training and test sets contained data from different hospitals in the CDR; this indicates that substantial distribution shifts can occur even in a relatively homogenous population in close geographical proximity using the same health care system and clinical guidelines.”
Quote in the paperSubstantial claim - Interpretationreasoning§ Key Points
Detecting progression to schizophrenia through machine learning may reduce diagnostic delay and duration of untreated illness.
“These findings suggest that detecting progression to schizophrenia through machine learning based on routine clinical data is feasible, which may reduce diagnostic delay and duration of untreated illness.”
Quote in the paperSubstantial claim - Interpretationown data§ Key Points
Detecting progression to schizophrenia through machine learning based on routine clinical data is feasible.
“These findings suggest that detecting progression to schizophrenia through machine learning based on routine clinical data is feasible, which may reduce diagnostic delay and duration of untreated illness.”
Quote in the paperSubstantial claim - Premisecitation§ Discussion
Most individuals with schizophrenia have the paranoid schizophrenia subtype (ICD-10 code F20), accounting for approximately 72% in the CDR.
“In contrast, most individuals with schizophrenia have the paranoid schizophrenia subtype (ICD-10 code F20), accounting for approximately 72% in the CDR,29 which is relatively more homogenous in its presentation.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [29] Köhler-Forsberg O, Antonsen S, Pedersen CB, Mortensen PB, McGrath JJ, Mors O. Schizophrenia spectrum disorders in Denmark between 2000 and 2018: Incidence and early diagnostic transition. Acta Psychiatr Scand. 2023;148(2):190-198. doi: 10.1111/acps.13565 [DOI] [PubMed] [Google Scholar]
Not checked in the free run.
- Interpretationown data§ Discussion
Substantial distribution shifts can occur even in a relatively homogenous population in close geographical proximity using the same health care system and clinical guidelines.
“The training and test sets contained data from different hospitals in the CDR; this indicates that substantial distribution shifts can occur even in a relatively homogenous population in close geographical proximity using the same health care system and clinical guidelines.”
Quote in the paperSubstantial claim - Interpretationown data§ Discussion
Model performance seems mainly to be driven by the inclusion of text-based predictors extracted from clinical notes.
“Model performance seems mainly to be driven by the inclusion of text-based predictors extracted from clinical notes.”
Quote in the paperSubstantial claim - Interpretationown data§ Results
It is possible to predict diagnostic transition to schizophrenia and bipolar disorder from routine clinical data extracted from EHRs.
“These findings suggest that it is possible to predict diagnostic transition to schizophrenia and bipolar disorder from routine clinical data extracted from EHRs, with schizophrenia being notably easier to predict than bipolar disorder.”
Quote in the paperSubstantial claim - Interpretationown data§ Discussion
Text models might be at risk of fitting to already present clinical suspicion and thereby provide less value.
“Second, text models might be at risk of fitting to already present clinical suspicion and thereby provide less value.”
Quote in the paperNot a substantial claim - Interpretationcitation§ Discussion
Performance from our models predicting schizophrenia is in line with the literature, for example with Irving et al, which achieved a Harrell C of 0.79 to 0.86 for 10-year survival prediction of onset of psychosis from an index date.
“Performance from our models predicting schizophrenia is in line with the literature, for example with Irving et al,12 which achieved a Harrell C of 0.79 to 0.86 for 10-year survival prediction of onset of psychosis from an index date.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [12] Irving J, Patel R, Oliver D, et al. Using natural language processing on electronic health records to enhance detection and prediction of psychosis risk. Schizophr Bull. 2021;47(2):405-414. doi: 10.1093/schbul/sbaa126 [DOI] [PMC free article] [PubMed] [Google Scholar] source
Not checked in the free run.
- Interpretationown data§ Conclusions
Text-based features from clinical notes show great promise for improving prediction of psychiatric outcomes.
“These findings suggest that text-based features from clinical notes show great promise for improving prediction of psychiatric outcomes.”
Quote in the paperSubstantial claim - Interpretationreasoning§ Discussion
The prediction models are primarily useful for patients who are progressing from less severe mental disorder to schizophrenia or bipolar disorder.
“Consequently, the prediction models are primarily useful for patients who are progressing from less severe mental disorder to schizophrenia or bipolar disorder.”
Quote in the paperSubstantial claim - Study design and conductown data§ Discussion
Our dataset only included data on medication administered during inpatient stays, not prescribed medication.
“Fifth, our dataset only included data on medication administered during inpatient stays, not prescribed medication.”
Quote in the paperSubstantial claim - Premisecitation§ Discussion
Wang et al achieved an AUROC of 0.80 in predicting early onset bipolar disorder 3 years into the future from a single randomly sampled time between age 10 to 25 years per patient.
“Wang et al13 achieved an AUROC of 0.80 in predicting early onset bipolar disorder 3 years into the future from a single randomly sampled time between age 10 to 25 years per patient.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [13] Wang B, Sheu YH, Lee H, Mealer RG, Castro VM, Smoller JW. Machine Learning models for the prediction of early-onset bipolar using electronic health records. . MedRxiv. Preprint posted online February 21, 2024. doi: 10.1101/2024.02.19.24302919 [DOI] [PMC free article] [PubMed] source
Not checked in the free run.
- Interpretationown data§ Discussion
Models predicting schizophrenia might be more suitable for implementation than those predicting bipolar disorder due to substantially better predictive performance.
“Models predicting schizophrenia might be more suitable for implementation than those predicting bipolar disorder due to substantially better predictive performance.”
Quote in the paperSubstantial claim - Interpretationreasoning§ Discussion
Direct comparison is difficult because most studies only make a single prediction at an index date or only report aggregated measures such as AUROC or the C-index.
“In summary, direct comparison is difficult because most studies only make a single prediction at an index date or only report aggregated measures such as AUROC or the C-index.”
Quote in the paperSubstantial claim - Interpretationnone given§ Discussion
The generalizability of the prediction models outside the Psychiatric Services in the Central Denmark Region is unknown.
“Fourth, the generalizability of the prediction models outside the Psychiatric Services in the Central Denmark Region is unknown.”
Quote in the paperNot a substantial claim - Interpretationreasoning§ Discussion
Patients whose initial contact to the psychiatric services is due to clinical suspicion of schizophrenia or bipolar disorder will not see additional benefits from the model.
“Patients whose initial contact to the psychiatric services is due to clinical suspicion of schizophrenia or bipolar disorder will not see additional benefits from the model.”
Quote in the paperSubstantial claim - Interpretationreasoning§ Discussion
Because most antipsychotic medication is prescribed rather than administered, we were unable to estimate the number of patients already receiving antipsychotic medication.
“Because most antipsychotic medication is prescribed (eg, to outpatients) rather than administered, we were unable to estimate the number of patients already receiving antipsychotic medication.”
Quote in the paperSubstantial claim- Rests on an assumption: Most antipsychotic medication is prescribed rather than administered.
- Interpretationreasoning§ Discussion
The study's focus on a high-risk population likely limits the generalizability of the findings to broader, low-risk populations such as those attending primary care.
“The study’s focus on a high-risk population likely limits the generalizability of the findings to broader, low-risk populations such as those attending primary care.”
Quote in the paperSubstantial claim
- Classified as probably not a substantial claim. 22
- May combine two claims. 18
Probably not a substantial claim: may not pass S3 (load-bearing).
Findingown data§ ResultsThe joint model performed slightly better on relatively young patients and older patients.
“The model performed slightly better on relatively young patients and older patients.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown data§ DiscussionA model predicting the progression to either schizophrenia or bipolar disorder achieved an AUROC of 0.64 on the test set, with notable disparity in predictive performance between the 2 disorders.
“A model predicting the progression to either of the 2 disorders achieved an AUROC of 0.64 on the test set, with notable disparity in predictive performance between the 2 disorders.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown data§ ResultsThe model for schizophrenia was highly robust across stratifications, with slightly better performance for older patients.
“eFigure 5 in Supplement 1 shows that the model for schizophrenia was highly robust across stratifications, with slightly better performance for older patients.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingnone given§ DiscussionAdmission and discharge directly pertained to hospitalization.
“Specifically, admission and discharge directly pertained to hospitalization.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingown data§ DiscussionThe terms play and the game often described patients' interactions with staff or other patients during their inpatient stay.
“The terms play and the game often described patients’ interactions with staff or other patients (eg, playing board games) during their inpatient stay.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Findingown data§ DiscussionSymptom-related terms included the voices, typically referring to auditory hallucinations, while female friends was often used to describe social interactions or lack thereof.
“Symptom-related terms included the voices, typically referring to auditory hallucinations, while female friends was often used to describe social interactions or lack thereof (ie, social withdrawal).”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown data§ ResultsThe study included 24,449 unique patients (median age 32.2 years; 56.6% female) and 398,922 outpatient contacts eligible for prediction.
“The cohort consisted of 24 449 unique patients (median [Q1-Q3] age at time of prediction, 32.2 [24.2-42.5] years; 13 843 female [56.6%]) with 398 922 outpatient contacts eligible for prediction.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingown data§ DiscussionThe phrase early morning frequently appeared at the beginning of notes documenting daily activities of inpatients.
“The phrase early morning frequently appeared at the beginning of notes documenting daily activities of inpatients.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingown data§ ResultsNo noticeable trends were observed in the performance of the joint model across calendar time.
“No noticeable trends were observed in the performance across calendar time.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingown data§ DiscussionThe term explain commonly appeared when patients struggled to articulate the reasons for their actions or experiences, potentially reflecting delusions or derealization.
“The term explain commonly appeared when patients struggled to articulate the reasons for their actions or experiences, potentially reflecting delusions or derealization.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown data§ DiscussionThe median time from the first positive prediction to the outcome was highest for the model using only structured predictors, despite having the lowest overall performance in terms of AUROC.
“As shown in eTable 6 in Supplement 1, the median time from the first positive prediction to the outcome was highest for the model using only structured predictors, despite having the lowest overall performance in terms of AUROC.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Premisecitation§ IntroductionMany patients who eventually receive a diagnosis of schizophrenia or bipolar disorder have previously received treatment for other and less severe mental disorders.
“In fact, many patients who eventually receive a diagnosis of schizophrenia or bipolar disorder have previously received treatment for other and less severe mental disorders.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [8] Musliner KL, Østergaard SD. Patterns and predictors of conversion to bipolar disorder in 91 587 individuals diagnosed with unipolar depression. Acta Psychiatr Scand. 2018;137(5):422-432. doi: 10.1111/acps.12869 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
- cites [9] Musliner KL, Munk-Olsen T, Mors O, Østergaard SD. Progression from unipolar depression to schizophrenia. Acta Psychiatr Scand. 2017;135(1):42-50. doi: 10.1111/acps.12663 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Premisecitation§ IntroductionBoth schizophrenia and bipolar disorder have severe negative consequences for social functioning, work ability, and lifespan.
“Indeed, both disorders have severe negative consequences for social functioning, work ability, and lifespan.”
Quote in the paperNot a substantial claim0 of 3 sources checked- cites [1] Goldman ML, Pincus HA, Mangurian C. Schizophrenia. N Engl J Med. 2020;382(6):583-584. doi: 10.1056/NEJMc1915943 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
- cites [2] Vieta E, Berk M, Schulze TG, et al. Bipolar disorders. Nat Rev Dis Primers. 2018;4:18008. doi: 10.1038/nrdp.2018.8 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
- cites [3] Laursen TM, Wahlbeck K, Hällgren J, et al. Life expectancy and death by diseases of the circulatory system in patients with bipolar disorder or schizophrenia in the Nordic countries. PLoS One. 2013;8(6):e67133. doi: 10.1371/journal.pone.0067133 [DOI] [PMC free article] [PubMed] [Google Scholar] source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Premisecitation§ IntroductionSchizophrenia and bipolar disorder are severe mental disorders that often impair the ability to lead a normal life.
“Schizophrenia and bipolar disorder are severe mental disorders that often impair the ability to lead a normal life.”
Quote in the paperNot a substantial claim0 of 2 sources checked- cites [1] Goldman ML, Pincus HA, Mangurian C. Schizophrenia. N Engl J Med. 2020;382(6):583-584. doi: 10.1056/NEJMc1915943 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
- cites [2] Vieta E, Berk M, Schulze TG, et al. Bipolar disorders. Nat Rev Dis Primers. 2018;4:18008. doi: 10.1038/nrdp.2018.8 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
May combine two claims.
Premisecitation§ IntroductionTimely diagnosis of schizophrenia and bipolar disorder is challenging due to the prodromal phase, in which patients do not yet meet full diagnostic criteria, and due to symptom overlap with other disorders such as anxiety and depression.
“However, timely diagnosis of schizophrenia and bipolar disorder is challenging due to the prodromal phase, in which patients do not yet meet full diagnostic criteria, and due to symptom overlap with other disorders such as anxiety and depression.”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [1] Goldman ML, Pincus HA, Mangurian C. Schizophrenia. N Engl J Med. 2020;382(6):583-584. doi: 10.1056/NEJMc1915943 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
- cites [6] Hafeman DM, Merranko J, Axelson D, et al. Toward the definition of a bipolar prodrome: dimensional predictors of bipolar spectrum disorders in at-risk youths. Am J Psychiatry. 2016;173(7):695-704. doi: 10.1176/appi.ajp.2015.15040414 [DOI] [PMC free article] [PubMed] [Google Scholar] source
Not checked in the free run.
- cites [7] Fusar-Poli P, Oliver D, Spada G, Estrade A, McGuire P. The case for improved transdiagnostic detection of first-episode psychosis: electronic health record cohort study. Schizophr Res. 2021;228:547-554. doi: 10.1016/j.schres.2020.11.031 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
May combine two claims.
Premisecitation§ MethodsSchizophrenia and bipolar disorder have low prevalence in younger individuals and heterogeneous symptoms in late-onset.
“Only patients aged 15 to 60 years were included due to the low prevalence of schizophrenia and bipolar disorder in younger individuals (eFigure 1 in Supplement 1) and heterogeneous symptoms in late-onset schizophrenia and bipolar disorder.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [20] Howard R, Rabins PV, Seeman MV, Jeste DV; The International Late-Onset Schizophrenia Group . Late-onset schizophrenia and very-late-onset schizophrenia-like psychosis: an international consensus. Am J Psychiatry. 2000;157(2):172-178. doi: 10.1176/appi.ajp.157.2.172 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
- cites [21] Schürhoff F, Bellivier F, Jouvent R, et al. Early and late onset bipolar disorders: two different forms of manic-depressive illness? J Affect Disord. 2000;58(3):215-221. doi: 10.1016/S0165-0327(99)00111-1 [DOI] [PubMed] [Google Scholar] source
Not checked in the free run.
Classified as probably not a substantial claim.
Study design and conductown data§ MethodsThe data were split into a training set and an external test set by hospital site, with western and eastern hospitals for training and central hospitals for testing.
“The data were split into a training set and an external test set by hospital site. Specifically, patient contacts with the hospitals in the western and eastern part of the CDR (Aarhus, Herning, Holstebro, Randers, Horsens, and Gødstrup) were used for training, while patient contacts with the central part of the region (Silkeborg and Viborg) were used for testing”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Study design and conductreasoning§ MethodsPredictions were issued at the earliest 3 months after a patient's first contact with psychiatric services to avoid flagging patients currently under assessment.
“Additionally, to avoid flagging patients currently under assessment for one of the disorders, predictions were issued at the earliest 3 months after a patient’s first contact with the psychiatric services in the CDR.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Study design and conductreasoning§ MethodsPrediction times occurring after November 21, 2016, were removed because they did not have the required 5 years of follow-up.
“Prediction times occurring after November 21, 2016, were removed because they did not have the required 5 years of follow-up”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductown data§ MethodsStratified analyses of the stability of model predictions in terms of AUROC over time, demographics, and individual outcomes were performed on the test set.
“We performed stratified analyses of the stability of model predictions in terms of AUROC over time, demographics, and individual outcomes (schizophrenia or bipolar disorder) on the test set.”
Quote in the paperNot a substantial claimMay combine two claims.
Interpretationown data§ DiscussionThe performance decline across sites was not attributable to hyperparameter overfitting, but instead results from distribution shifts between sites.
“This performance closely mirrors the out-of-fold performance observed in the training set of the models presented in the results section, which suggests that the performance decline across sites was not attributable to hyperparameter overfitting, but instead results from distribution shifts between sites.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone given§ MethodsA model using the best performing hyperparameters was trained on an 85% to 15% train-test split, stratified only by patient, not hospital site, to assess performance on unseen patients attending the same psychiatric services.
“Additionally, a model using the best performing hyperparameters was trained on an 85% to 15% train-test split, stratified only by patient, not hospital site, to assess performance on unseen patients attending the same psychiatric services.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductown data§ MethodsAs a sensitivity analysis, the best-performing joint model was tested on predicting schizophrenia or bipolar disorder separately.
“As a sensitivity analysis of the best-performing joint model, we tested how it performed in predicting schizophrenia or bipolar disorder separately.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim. May combine two claims.
Study design and conductnone given§ MethodsOnly routine clinical data from EHRs were considered for predictors; there was no data collection for the purpose of this study.
“Notably, only routine clinical data from the EHRs were considered for predictors. There was no data collection for the purpose of this study.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim. May combine two claims.
Study design and conductnone given§ MethodsThe best-performing model was retrained on the entire training set and applied to the test set.
“The best-performing model in terms of the highest area under the receiver operating characteristic curve (AUROC) after hyperparameter tuning was retrained on the entire training set and applied to the test set”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Study design and conductnone given§ MethodsOutpatient contacts occurring after the diagnosis of schizophrenia or bipolar disorder were removed.
“Outpatient contacts occurring after the diagnosis of schizophrenia or bipolar disorder were removed.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationreasoning§ DiscussionThe discrepancy in predictive performance between schizophrenia and bipolar disorder may be attributed to the relatively larger heterogeneity within bipolar disorder compared with schizophrenia and the distinctiveness of the psychotic symptoms of schizophrenia.
“This discrepancy may be attributed to the relatively larger heterogeneity within bipolar disorder compared with schizophrenia and the distinctiveness of the psychotic symptoms of schizophrenia.”
Quote in the paperNot a substantial claim- Rests on an assumption: Bipolar disorder covers a wide spectrum of illness manifestations, with some individuals initially presenting with mania and others with depression.
May combine two claims.
Interpretationreasoning§ ConclusionsThe model predicting schizophrenia performed substantially better than the model predicting bipolar disorder, likely due to heterogenic clinical manifestations of the latter.
“The model predicting schizophrenia performed substantially better than the model predicting bipolar disorder, likely due to heterogenic clinical manifestations of the latter.”
Quote in the paperSubstantial claim- Rests on an assumption: Bipolar disorder covers a wide spectrum of illness manifestations, with some individuals initially presenting with mania and others with depression.
May combine two claims.
Interpretationown data§ DiscussionWhile text models might lead to more cases being identified correctly, this might be at the cost of less lead time.
“This finding indicates that while text models might lead to more cases being identified correctly, this might be at the cost of less lead time.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.
Interpretationnone given§ DiscussionRetraining would likely be required for successful transfer of the models, especially if testing where a different language is used.
“Likely, retraining would be required for successful transfer, especially if testing the model where a different language is used.”
Quote in the paperNot a substantial claimMay combine two claims.
Premisecitation§ DiscussionHarrell C and AUROC are equivalent in binary outcomes, but direct comparisons cannot be made with censored data such as those used by Irving and colleagues.
“Harrell C and AUROC are equivalent in binary outcomes, but direct comparisons cannot be made with censored data such as those used by Irving and colleagues.12”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [12] Irving J, Patel R, Oliver D, et al. Using natural language processing on electronic health records to enhance detection and prediction of psychosis risk. Schizophr Bull. 2021;47(2):405-414. doi: 10.1093/schbul/sbaa126 [DOI] [PMC free article] [PubMed] [Google Scholar] source
Not checked in the free run.
May combine two claims.
Study design and conductnone given§ DiscussionThe data were restricted to patients receiving psychiatric treatment and did not contain information from primary care.
“The data were restricted to patients receiving psychiatric treatment and did not contain information from primary care.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone given§ DiscussionThe test performance of our models was evaluated on data from 2 hospital sites.
“Third, the test performance of our models was evaluated on data from 2 hospital sites.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductreasoning§ DiscussionComputational limitations precluded internal-external cross-validation because analyses were restricted to in-house hardware due to the highly sensitive nature of the EHR data.
“Computational limitations, however, precluded this option because we were restricted to run our analyses on in-house hardware (on the CDR server) due to the highly sensitive nature of the EHR data.”
Quote in the paperNot a substantial claim
The paper relies on this without support.
The diagnostic codes in the EHR accurately reflect the clinical diagnoses of schizophrenia and bipolar disorder.
“Diagnostic progression to schizophrenia was defined as the time of the first International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10) diagnosis (codes in parentheses) of either schizophrenia (F20) or schizoaffective disorder (F25).”
The paper relies on this without support.
The 5-year look-ahead window is sufficient to capture diagnostic progression.
“Models were trained to predict the incidence of the outcomes 5 years into the future (5-year look-ahead window).”
The paper relies on this without support.
The hospital site split provides a valid external test set representative of the overall population.
“The data were split into a training set and an external test set by hospital site.”
The paper relies on this without support.
The synthetic data augmentation does not introduce bias that affects model performance.
“Data augmentation using synthetically generated data has been proposed to improve performance on multiple classification tasks within health care.26,27 During training, experiments were therefore conducted to augment the training data with synthetic data generated using 2 methods: tabular denoising diffusion probabilistic models (TabDDPM) and synthetic minority oversampling technique.”
The paper relies on this without support.
The text embeddings capture clinically relevant information from clinical notes.
“Free-form clinical notes (note types are specified in eTable 2 in Supplement 1) from the EHRs were embedded as numerical feature vectors using 3 different methods: (1) term-frequency of predefined words describing psychopathology, (2) term frequency-inverse document frequency (TF-IDF), and (3) sentence transformers.23”
The paper relies on this without support.
Bipolar disorder covers a wide spectrum of illness manifestations, with some individuals initially presenting with mania and others with depression.
“Bipolar disorder covers a wide spectrum of illness manifestations, with some individuals initially presenting with mania and others with depression.”
The paper relies on this without support.
Most antipsychotic medication is prescribed rather than administered.
“Because most antipsychotic medication is prescribed (eg, to outpatients) rather than administered, we were unable to estimate the number of patients already receiving antipsychotic medication.”
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
Machine learning approaches have previously been utilized to predict adverse mental health outcomes, including progression of schizophrenia and bipolar disorder and involuntary psychiatric admission.
“Machine learning approaches have previously been utilized to predict adverse mental health outcomes, including progression of schizophrenia and bipolar disorder [51] and involuntary psychiatric admission [52].”
From Factors associated with postpartum depression symptoms following antepartum hospitalization (2026) · cites this paper as [51]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Software citation was accessed before its stated update
MinorReference 28 says the GitHub resource was updated in January 2025 but accessed on January 28, 2024. The access date precedes the stated update by about a year.
“Updated January 2025.”
“Accessed January 28, 2024.”
Cites
XGBoost: A Scalable Tree Boosting System
Chen, T., & Guestrin, C. (2016) · Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining · doi:10.1145/2939672.2939785 · Cited by 54,569 · see citing works
Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks
Reimers, N., & Gurevych, I. (2019) · Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) · doi:10.18653/v1/d19-1410 · Cited by 12,778 · see citing works
Larian Studios Dev Team (2024) · doi:10.1145/3292500.3330701 · Cited by 8,618 · see citing works
Collins, G. S., Moons, K. G. M., Dhiman, P., Riley, R. D., Beam, A. L., Van Calster, B., Ghassemi, M., Liu, X., Reitsma, J. B., van Smeden, M., et al. (2024) · BMJ · doi:10.1136/bmj-2023-078378 · Cited by 4,718 · see citing works
Synthetic data augmentation using GAN for improved liver lesion classification
Frid-Adar, M., Klang, E., Amitai, M. M., Goldberger, J., & Greenspan, H. K. (2018) · IEEE International Symposium on Biomedical Imaging (ISBI) · doi:10.1109/isbi.2018.8363576 · Cited by 919 · see citing works
Vieta, E., Berk, M., Schulze, T. G., Carvalho, A. F., Suppes, T., Calabrese, J. R., Gao, K., Miskowiak, K. W., & Grande, I. (2018) · Nature Reviews Disease Primers · doi:10.1038/nrdp.2018.8 · Cited by 891 · see citing works
Machine Learning for Precision Psychiatry: Opportunities and Challenges
Bzdok, D., & Meyer‐Lindenberg, A. (2017) · Biological Psychiatry Cognitive Neuroscience and Neuroimaging · doi:10.1016/j.bpsc.2017.11.007 · Cited by 833 · see citing works
Evaluation of clinical prediction models (part 1): from development to external validation
Collins, G. S., Dhiman, P., Ma, J., Schlussel, M. M., Archer, L., Van Calster, B., Harrell, F. E., Martin, G. P., Moons, K. G. M., van Smeden, M., et al. (2024) · BMJ · doi:10.1136/bmj-2023-074819 · Cited by 798 · see citing works
Penttilä, M., Jääskeläinen, E., Hirvonen, N., Isohanni, M., & Miettunen, J. (2014) · British Journal of Psychiatry · doi:10.1192/bjp.bp.113.127753 · Cited by 716 · see citing works
Howard, R., Rabins, P. V., Seeman, M. V., Jeste, D. V., & Late-Onset, T. I. (2000) · American Journal of Psychiatry · doi:10.1176/appi.ajp.157.2.172 · Cited by 678 · see citing works
Natural language processing applied to mental illness detection: a narrative review
Zhang, T., Schoene, A. M., Ji, S., & Ananiadou, S. (2022) · npj Digital Medicine · doi:10.1038/s41746-022-00589-7 · Cited by 476 · see citing works
Laursen, T. M., Wahlbeck, K., Hällgren, J., Westman, J., Ösby, U., Alinaghizadeh, H., Gissler, M., & Nordentoft, M. (2013) · PLoS ONE · doi:10.1371/journal.pone.0067133 · Cited by 313 · see citing works
Early and late onset bipolar disorders: two different forms of manic-depressive illness?
Schürhoff, F., Bellivier, F., Jouvent, R., Mouren-Siméoni, M.-C., Bouvard, M., Allilaire, J.-F., & Leboyer, M. (2000) · Journal of Affective Disorders · doi:10.1016/s0165-0327(99)00111-1 · Cited by 297 · see citing works
Hafeman, D. M., Merranko, J., Axelson, D., Goldstein, B. I., Goldstein, T., Monk, K., Hickey, M. B., Sakolsky, D., Diler, R., Iyengar, S., et al. (2016) · American Journal of Psychiatry · doi:10.1176/appi.ajp.2015.15040414 · Cited by 212 · see citing works
Rumshisky, A., Ghassemi, M., Naumann, T., Peter Szolovits, B., Castro, V. M., McCoy, T. H., & Perlis, R. H. (2016) · Translational Psychiatry · doi:10.1038/tp.2015.182 · Cited by 190 · see citing works
Altamura, A. C., Buoli, M., Caldiroli, A., Caron, L., Cumerlato Melter, C., Dobrea, C., Cigliobianco, M., & Zanelli Quarantini, F. (2015) · Journal of Affective Disorders · doi:10.1016/j.jad.2015.04.024 · Cited by 162 · see citing works
Irving, J., Patel, R., Oliver, D., Colling, C., Pritchard, M., Broadbent, M., Baldwin, H., Stahl, D., Stewart, R., & Fusar-Poli, P. (2021) · Schizophrenia Bulletin · doi:10.1093/schbul/sbaa126 · Cited by 92 · see citing works
Musliner, K. L., & Østergaard, S. D. (2018) · Acta Psychiatrica Scandinavica · doi:10.1111/acps.12869 · Cited by 79 · see citing works
Self-citationDebray, T. P. A., Collins, P. G. S., Riley, R. D., Snell, K. I. E., CALSTER, B. V., Reitsma, J. B., & Moons, K. G. M. (2023) · BMJ · doi:10.1136/bmj-2022-071058 · Cited by 62 · see citing works
Raket, L. L., Jaskolowski, J., Kinon, B. J., Brasen, J. C., Jönsson, L., Wehnert, A., & Fusar-Poli, P. (2020) · The Lancet Digital Health · doi:10.1016/s2589-7500(20)30024-8 · Cited by 60 · see citing works
Bernstorff, M., Hansen, L., Enevoldsen, K., Damgaard, J. G., Hæstrup, F., Perfalk, E., Danielsen, A. A., & Østergaard, S. D. (2024) · Acta Psychiatrica Scandinavica · doi:10.1111/acps.13687 · Cited by 22 · see citing works
Self-citationProgression from unipolar depression to schizophrenia.
Musliner, K. L., Munk‐Olsen, T., Mors, O., & Østergaard, S. D. (2017) · Acta Psychiatrica Scandinavica · doi:10.1111/acps.12663 · Cited by 21 · see citing works
Self-citationBernstorff, M., Hansen, L., Perfalk, E., Danielsen, A. A., & Østergaard, S. D. (2022) · Acta Psychiatrica Scandinavica · doi:10.1111/acps.13463 · Cited by 18 · see citing works
Self-citationHansen, L., Enevoldsen, K., Bernstorff, M., Nielbo, K. L., Danielsen, A. A., & Østergaard, S. D. (2021) · Acta Neuropsychiatrica · doi:10.1017/neu.2021.22 · Cited by 17 · see citing works
Self-citationLexical stability of psychiatric clinical notes from electronic health records over a decade
Hansen, L., Enevoldsen, K., Bernstorff, M., Perfalk, E., Danielsen, A. A., Nielbo, K. L., & Østergaard, S. D. (2023) · Acta Neuropsychiatrica · doi:10.1017/neu.2023.46 · Cited by 14 · see citing works
Self-citationGoldman, M. L., Pincus, H. A., & Mangurian, C. (2020) · New England Journal of Medicine · doi:10.1056/nejmc1915943 · Cited by 13 · see citing works
Köhler‐Forsberg, O., Antonsen, S., Pedersen, C. B., Mortensen, P. B., McGrath, J. J., & Mors, O. N. (2023) · Acta Psychiatrica Scandinavica · doi:10.1111/acps.13565 · Cited by 12 · see citing works
Timeseriesflattener: A Python package for summarizing features from (medical) time series.
Bernstorff, M., Enevoldsen, K., Damgaard, J., Danielsen, A., & Hansen, L. (2023) · Journal of Open Source Software · doi:10.21105/joss.05197 · Cited by 10 · see citing works
Fusar-Poli, P., Oliver, D., Spada, G., Estrade, A., & McGuire, P. (2021) · Schizophrenia Research · doi:10.1016/j.schres.2020.11.031 · Cited by 6 · see citing works
SurvTimeSurvival: Survival Analysis On The Patient With Multiple Visits/RecordsPreprint
Lê, H., Eng-Jon, O., & Miroslaw, B. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2311.09854 · Cited by 1 · see citing works
Machine Learning Models for the Prediction of Early-Onset Bipolar Using Electronic Health RecordsPreprint
Wang, B., Sheu, Y.-H., Lee, H., Mealer, R. G., Castro, V. M., & Smoller, J. W. (2024) · medRxiv · doi:10.1101/2024.02.19.24302919 · Cited by 1 · see citing works
Flagged first, then by citations
Cited by
Sorted by citations
Rabbani, S. A., El‐Tanani, M., Sharma, S., Rabbani, S. S., El‐Tanani, Y., Kumar, R. R., & Saini, M. (2025) · BioMedInformatics · doi:10.3390/biomedinformatics5030037 · Cited by 102 · see citing works
“An XGBoost model applied to EHRs from 24,449 patients predicted diagnostic progression with an overall AUROC of 0.64–0.80 for schizophrenia and 0.62 for bipolar disorder, leveraging clinical notes to boost early detection [35].”
InfluentialStein, F., & Kircher, T. T. J. (2025) · Journal of Affective Disorders · doi:10.1016/j.jad.2025.119464 · Cited by 21 · see citing works
Cellat, E., & Demir, E. (2026) · Asian Journal of Psychiatry · doi:10.1016/j.ajp.2026.104828 · Cited by 9 · see citing works
Arribas, M., Micheli, A. D., Krakowski, K., Stahl, D. R., Correll, C. U., Young, A. H., Andreassen, O. A., Vieta, E., Arango, C., McGuire, P. K., et al. (2025) · The Lancet Psychiatry · doi:10.1016/s2215-0366(25)00307-4 · Cited by 8 · see citing works
The Use of Artificial Intelligence for Personalized Treatment in Psychiatry
Jalali, S., You, Q., Xu, V., Chen, L., Zini, J., Zamorano, T., Luo, Y., Friesen, T., Fontana, G. J., & Benrimoh, D. A. (2025) · Current Psychiatry Reports · doi:10.1007/s11920-025-01656-y · Cited by 8 · see citing works
“Current Psychiatry Reports (2026) 28:7 models on EHR data to predict 5-year transition to schizo - phrenia or bipolar disorder [43 ].”
Luo, X., Dong, J., & Li, T. (2025) · Inflammation · doi:10.1007/s10753-025-02344-y · Cited by 4 · see citing works
Prevalence estimates of nonordinary experiences depend on validity and measurement choices
Fischer, R. D., Bortolini, G. N., Bortolini, T., Maraldi, E. D. O., Hartle, L., Varejão, M. V. D. L., Laport, M. C., Granjeiro, G. F., de Oliveira, M. E. G., Ihm, E. D., et al. (2026) · Communications Psychology · doi:10.21203/rs.3.rs-6907742/v1 · Cited by 3 · see citing works
Sun, S., Wang, S., Gao, Y., & Cui, M. (2025) · Neuroscience · doi:10.1016/j.neuroscience.2025.05.007 · Cited by 2 · see citing works
Lewin, G., Abakasanga, E., Titcombe, I., Cosma, G., & Gangadharan, S. K. (2025) · BJPsych Advances · doi:10.1192/bja.2025.10133 · Cited by 2 · see citing works
“Hansen et al (2025) conducted a machine learning analysis of electronic health records from 24 449 psychiatric patients, demonstrating that diagnostic progression to schizophrenia or bipolar disorder could be predicted using routine clinical data, although with higher accuracy for the former than…”
Enhancing the early diagnosis of bipolar disorder with machine learning models
Alharithi, F. S. (2026) · Journal of Radiation Research and Applied Sciences · doi:10.1016/j.jrras.2026.102245 · Cited by 2 · see citing works
Blocking the Reflection: Milestones and Hurdles for Digital Twins in Mental Health
Verhees, F. G., Wiest, I. C., Kather, J. N., Kambeitz, J., & Mikoláš, P. (2026) · Pharmacopsychiatry · doi:10.1055/a-2816-2869 · Cited by 2 · see citing works
RETRACTED ARTICLE: MRI feature engineering and SVM framework for schizophrenia recognition
Liu, J., Liu, L., Wu, Y., Wang, Z., & Li, X. (2025) · Disability and Rehabilitation Assistive Technology · doi:10.1080/17483107.2025.2569801 · Cited by 1 · see citing works
RetractedCrocamo, C. (2026) · BJPsych Advances · doi:10.1192/bja.2025.10191 · Cited by 1 · see citing works
“In recent years, AI has gained increasing prominence in psychiatry, offering powerful tools to address digital health challenges such as early detection, risk prediction, treatment response and personalisation (Sheu 2023; Hansen 2025).”
Marafino, B. J., Kline‐Simon, A. H., Stavers-Sosa, I., Cronkite, D., Gerstley, L. D., Durojaiye, C., Kelley, A. M., Kiel, L. M., Ramaprasan, A., Carrell, D., et al. (2026) · BMC Psychiatry · doi:10.1186/s12888-026-07846-z · Cited by 2 · see citing works
“In a more recent study using EHR data from Danish psychiatric clinics, researchers sought to predict 5-year transition risk to schizophrenia or bipolar disorder by developing and externally testing a model that integrated medication lists, diagnoses, and the text of clinical notes [12].”
Hu, H., Zu, Y., Zhao, L., Zhang, Q., Zhou, G., Xu, P., Zhao, A., Yin, F., Sharma, L., & Chang, D. (2026) · Respiratory Research · doi:10.1186/s12931-026-03581-x · Cited by 1 · see citing works
“Although the PPV was relatively low, which is typical for low-incidence outcomes, the model ultimately provides a critical tool for aiding clinical decision-making [33].”
Yang, H., Chang, F., Muroi, F., Liu, Z., Zhang, W., & Cai, J. (2026) · Translational Psychiatry · doi:10.1038/s41398-026-03872-3 · Cited by 2 · see citing works
Artificial Intelligence–Empowered Multimodal Learning in Psychiatry: A Scoping Review
Li, F.-L., Li, P.-Z., Bhandarkar, A., Huo, N., Raja, K., Hong, N., Kumar, R., Altalib, H. H., Lu, Y.-W., Wang, T.-Y., et al. (2026) · Biological Psychiatry: Cognitive Neuroscience and Neuroimaging · doi:10.1016/j.bpsc.2026.03.013 · Cited by 1 · see citing works
Özsoy, F., Tasci, G., Taşçı, B., Doğan, Ş., & Tuncer, T. (2026) · World Journal of Psychiatry · doi:10.5498/wjp.v16.i5.116452 · Cited by 1 · see citing works
Guo, L., Zhang, Q., Sun, L., Hu, P., Wang, Z., Gu, J., Sun, Q., & He, Y. (2026) · Frontiers in Public Health · doi:10.3389/fpubh.2026.1914779 · Cited by 1 · see citing works
“Advanced algorithms, such as Extreme Gradient Boosting (XGBoost), can process high-dimensional interactions to forecast outcomes with superior accuracy (13, 14).”
Mwangi, B., Hamoudi, H. J. A. S., Sanches, M., Doğan, N., Chaudhary, P., Wu, M., Zunta‐Soares, G. B., Soares, J. C., Martin, A., & Soutullo, C. A. (2026) · npj Mental Health Research · doi:10.1038/s44184-026-00245-y · Cited by 1 · see citing works
Srivastava, K., & Srivastava, R. (2026) · Frontiers in Systems Biology · doi:10.3389/fsysb.2026.1898853 · Cited by 1 · see citing works
Hansen, L., Damgaard, J. G., Lundin, R. M., Danielsen, A. A., & Østergaard, S. D. (2025) · medRxiv · doi:10.1101/2025.06.19.25329678 · Cited by 1 · see citing works
Self-citationFactors associated with postpartum depression symptoms following antepartum hospitalization
Goulding, A. N., Palacios, D., Aras, S., Chen, H., Pasupuleti, S., Toscano, M., Cirino, N., Christie, I. C., Liu, Z., Miller, E. S., et al. (2026) · Pregnancy · doi:10.1002/pmf2.70319
Smucny, J. (2025) · Biological Psychiatry Cognitive Neuroscience and Neuroimaging · doi:10.1016/j.bpsc.2025.06.004
Kolding, S., Damgaard, J. G., Bernstorff, M., Hansen, L., Østergaard, S. D., & Danielsen, A. A. (2026) · Nordic Journal of Psychiatry · doi:10.1080/08039488.2026.2716241
“Patient assignment to data splits was determined using an a priori stratification based on multiple outcomes defined by the author group, ensuring consistent train–test splits across related studies [26–30].”
InfluentialSelf-citationKumar, S., Kumari, D., Panwar, A., Sagar, S., Herout, L., Namazi, H., & Bhati, N. S. (2026) · Scientific Reports · doi:10.1038/s41598-026-39104-3
“Nonetheless, the practicality for numerous present approaches is restricted by enduring problems like class imbalance, data sparsity, and model generalizability 15–17 .”
InfluentialFerro, E., Castano-Villegas, N., Cárdenas, M., Gómez, A., Torres-Delgado, C., Calderon, L. O., Barrientos, K. M., & Zea, J. (2026) · medRxiv · doi:10.64898/2026.05.07.26352679
Early Diagnosis of the Risk of Schizophrenia using Hybrid Machine Learning
Dahal, S., & N.Sivakumar, S. G. (2026) · International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV) · doi:10.1109/icicv68925.2026.11554750
Zhu, G., 周毓青, & Li, B. (2026) · Psychiatry Research · doi:10.1016/j.psychres.2026.117335
Artificial intelligence analytics for diagnosis, prognosis, and outcomes prediction
Jalali, S., You, Q., Xu, V., Chen, L., Zini, J., Zamorano, T., Luo, Y., Friesen, T., Fontana, G. J., & Benrimoh, D. A. (2026) · Elsevier eBooks · doi:10.1016/b978-0-443-33151-0.00009-5
Mapping GWAS-derived Genetic Variants to Mental Illness via Machine Learning
Hamssini, K. P., Katakam, K., B., K. M., Palak, Rashmi, K. B., & Shobana, T. S. (2026) · IEEE Guwahati Subsection Conference (GCON) · doi:10.1109/gcon69192.2026.11648910
“Other studies have incorporated imaging and genomic data to achieve multi-modal deep learning-based classification of schizophrenia [10] and prediction of diagnostic transition to schizophrenia or bipolar disorder using predictive machine learning models [11].”
Perfalk, E., Damgaard, J. G., Danielsen, A. A., & Østergaard, S. D. (2026) · Schizophrenia Research · doi:10.1016/j.schres.2026.08.029
Self-citationQin, L., Hong, J., Liu, J., Zhou, S., 刘金晶, Cao, W., Wu, Q., Zeng, X., & Li, T. (2026) · BMJ Open · doi:10.1136/bmjopen-2026-121354
“19 22 To enhance model efficiency and interpretability, feature selection was performed using the Boruta algorithm, which identifies all relevant features by comparing them with permuted shadow features.”
Digital screening tools for individuals at risk of mental disorders: a systematic review (Preprint)Preprint
Logeswaran, Y., La‐Torraca‐Vittori, P., Ahmed, M. W., Arribas, M., Wallman, P., Reilly, T. J., Capon, W., Aronica, R., Brodeur, S., Dusara, H., et al. (2026) · doi:10.2196/preprints.114283
- Authors
- Hansen, L., Bernstorff, M., Enevoldsen, K., Kolding, S., Damgaard, J. G., Perfalk, E., Nielbo, K. L., Danielsen, A. A., & Østergaard, S. D.
- Affiliations
- Aarhus University (DK)
- ORCID
- Venue
- JAMA Psychiatry
- Publisher
- American Medical Association (AMA)
- Published
- 2025-05-01
- Volume
- 82
- Issue
- 5
- Pages
- 459
- ISSN
- 2168-622X (ISSN-L)
- PMC
- PMC12551142
- OpenAlex
- W4407722299
- Cited by
- 34
- References
- 30
- Citations by year
- 2026: 24 · 2025: 10
- FWCI
- 13.64
- Type
- journal article
- Language
- en
- Open access
- green
- pmc.ncbi.nlm.nih.gov
- Locations
- JAMA Psychiatry (publishedVersion)
- PubMed (publishedVersion)
- PubMed Central (submittedVersion)
- Scopus (Elsevier) (submittedVersion)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
- Health Sciences › Medicine › Psychiatry and Mental health › Schizophrenia research and treatment
- Health Sciences › Medicine › Psychiatry and Mental health › Bipolar Disorder and Treatment
- Keywords
- schizophrenia, bipolar disorder, electronic health records, XGBoost, elastic net logistic regression, psychiatric outpatients, clinical notes, drug history, clinical diagnoses, delayed diagnosis, psychiatric comorbidity, Central Denmark Region, machine learning prediction models, clinical decision support, treatment initiation timing
- MeSH terms
- Machine Learning, Adolescent, Adult, Bipolar Disorder, Denmark, Female, Humans, Male, Middle Aged, Schizophrenia, Cohort Studies, Disease Progression, Young Adult, Electronic Health Records
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.
Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
