CoChat
Sign Up
Cited by 47Mental Health Research TopicsDigital Mental Health InterventionsEmotion and Mood Recognition

Personalised depression forecasting using mobile sensor data and ecological momentary assessment

Open access
Kathan, Alexander, Harrer, Mathias, Küster, Ludwig, Triantafyllopoulos, Andreas, He, Xiangheng, Milling, Manuel, Gerczuk, Maurice, Yan, Tianhao +5 more·2022·Frontiers in Digital Health
DOI 10.3389/fdgth.2022.964582PMC PMC9715619
Open this paper in CoChat — free accountOpen PDF

Abstract

Introduction: Digital health interventions are an effective way to treat depression, but it is still largely unclear how patients’ individual symptoms evolve dynamically during such treatments. Data-driven forecasts of depressive symptoms would allow to greatly improve the personalisation of treatments. In current forecasting approaches, models are often trained on an entire population, resulting in a general model that works overall, but does not translate well to each individual in clinically heterogeneous, real-world populations. Model fairness across patient subgroups is also frequently overlooked. Personalised models tailored to the individual patient may therefore be promising. Methods: We investigate different personalisation strategies using transfer learning, subgroup models, as well as subject-dependent standardisation on a newly-collected, longitudinal dataset of depression patients undergoing treatment with a digital intervention (N=65 patients recruited). Both passive mobile sensor data as well as ecological momentary assessments were available for modelling. We evaluated the models’ ability to predict symptoms of depression (Patient Health Questionnaire-2; PHQ-2) at the end of each day, and to forecast symptoms of the next day. Results: In our experiments, we achieve a best mean-absolute-error (MAE) of 0.801 (25% improvement) for predicting PHQ-2 values at the end of the day with subject-dependent standardisation compared to a non-personalised baseline (MAE=1.062). For one day ahead-forecasting, we can improve the baseline of 1.539 by 12% to a MAE of 1.349 using a transfer learning approach with shared common layers. In addition, personalisation leads to fairer models at group-level. Discussion: Our results suggest that personalisation using subject-dependent standardisation and transfer learning can improve predictions and forecasts, respectively, of depressive symptoms in participants of a digital depression intervention. We discuss technical and clinical limitations of this approach, avenues for future investigations, and how personalised machine learning architectures may be implemented to improve existing digital interventions for depression.

AI snapshot

Written from the paper’s full text.

TL;DR

This paper tests personalisation strategies for forecasting daily depressive symptoms from mobile sensor and ecological momentary assessment data in patients undergoing a digital depression intervention. It finds that subject-dependent standardisation improves same-day PHQ-2 prediction and transfer learning improves next-day forecasting, while personalisation also changes group- and individual-level fairness.

Key findings

  • For end-of-day PHQ-2 prediction, subject-dependent standardisation achieved the best MAE of 0.801 and Spearman’s ρ = 0.728, compared with the non-personalised baseline MAE = 1.062 and ρ = 0.604; the paper describes this as a 25% improvement.
  • For one-day-ahead PHQ-2 forecasting, transfer learning with shared common layers achieved the best MAE of 1.349 and ρ = 0.349, improving the baseline MAE = 1.539 and ρ = 0.105 by 12%.
  • Group-level sex fairness improved with personalisation: the baseline showed negative sex fairness scores and biases, while subgroup models nearly eliminated prediction bias and transfer learning performed best for forecasting bias.
  • Individual-level fairness was mixed: personalisation increased the Gini Coefficient in most cases, but DMLI was consistently higher and DMUI consistently lower, meaning “winners” diverged further from the median while “losers” moved closer to it.

Methods

  • The MAIKI dataset was collected in a prospectively registered feasibility trial; N=65 patients with elevated depressive symptoms were recruited, received an 8-week digital depression intervention, and provided active EMA plus passive mobile sensor/location data. After pre-defined data-quality criteria, analyses used 16 quality-assured participants.
  • The baseline model was a GRU-based recurrent neural network using data up to 7 days before the target PHQ-2 label. Three personalisation strategies were tested: transfer learning with shared common layers, gender-based subgroup models, and subject-dependent standardisation.
  • Evaluation used 3-fold cross-validation with subject-dependent splits, MAE and Spearman’s ρ for prediction/forecasting, plus group-level sex fairness scores and proposed individual-level fairness indices (Gini Coefficient, DMLI, DMUI).

Limitations

  • The paper states that missing values meant only a subset of MAIKI patients could be included; seamless passive sensor recording was technically challenging because patients used privately owned smartphones, and some reported difficulties installing/navigating the companion app.
  • The paper states that most important features were actively assessed EMA data, and that some personalised models require several weeks of individual training data, creating a “burn-in” constraint and risk of overfitting if PHQ values have limited variance.
  • Evident from the text: the quality-assured analysis sample was small (N=16) and predominantly female (81.2% of quality-assured participants), and group-level fairness could only be assessed by biological sex because it was the only relevant group-level variable available.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3389/fdgth.2022.964582

Cites

91 of 91 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 46 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Transfer learning with shared common layers improved one-day-ahead forecasting from a baseline MAE of 1.539 by 12% to a MAE of 1.349.

    “For one day ahead-forecasting, we can improve the baseline of 1.539 by 12% to a MAE of 1.349 using a transfer learning approach with shared common layers.”
    Quote in the paperSubstantial claim
  • Findingown data

    Subject-dependent standardisation achieved a best mean-absolute-error of 0.801 (25% improvement) for predicting PHQ-2 values at the end of the day compared to a non-personalised baseline MAE of 1.062.

    “In our experiments, we achieve a best mean-absolute-error (MAE) of 0.801 (25% improvement) for predicting PHQ-2 values at the end of the day with subject-dependent standardisation compared to a non-personalised baseline (MAE=1.062).”
    Quote in the paperSubstantial claim
  • Findingown data

    The random forest analysis indicated that EMA data in particular make a considerable contribution to the PHQ-2 prediction result.

    “The analysis indicates that the EMA data in particular make a considerable contribution to the prediction result.”
    Quote in the paperNot a substantial claim
  • Findingnone given

    A considerable percentage of the study participants only provided a small amount of data.

    “However, a considerable percentage of the study participants only provided a small amount of data.”
    Quote in the paperNot a substantial claim
  • Findingown data

    In the subset of patients with sufficient active EMA and passive mobile sensor data, all investigated personalisation strategies improved end-of-day PHQ-2 prediction compared to a general RNN baseline.

    “In the case of PHQ-2 prediction, the personalised models clearly outperform the non-personalised baseline.”
    • Also stated in: § Discussion · plain
      “In a subset of patients who provided sufficient active EMA as well as passive mobile sensor data, we found that all investigated personalisation strategies lead to improved predictions of the end-of-day depressive symptom severity (MAE=0.801 to 0.990), compared to a general RNN model (MAE=1.062).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The subset of patients with sufficient active EMA and passive mobile sensor data is representative of the broader patient population with elevated depressive symptoms.
  • Classified as probably not a substantial claim. 32
  • May combine two claims. 27
  • The quote only partly states this claim. 2

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    Personalisation leads to fairer models at group-level.

    “In addition, personalisation leads to fairer models at group-level.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Subject-dependent standardisation yields the best PHQ-2 prediction result, with MAE .801 and Spearman's ρ 0.728, compared to the non-personalised baseline MAE 1.062 and ρ .604.

    “The best result is obtained using subject-dependent standardisation which yields a MAE of .801 and Spearman’s ρ correlation of 0.728 compared to the non personalised model that achieves 1.062 and .604, respectively.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    MAIKI is a real-world dataset and therefore contains missing data.

    “As described in section 2.2 MAIKI is a real-world dataset and therefore contains missing data.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    A total of N=65 patients were recruited for the study between May 2021 and September 2021.

    “Between May 2021 and September 2021, a total of N=65 patients were recruited for the study.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Sleep time as part of user activity features and several GPS features are among the most relevant features for PHQ-2 prediction.

    “In addition, sleep time as part of the user activity features as well as several GPS features are among the most relevant features.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Personalised transfer learning with shared common backbone and personalised layers yields the best one-day-ahead PHQ-2 forecasting result, with MAE 1.349 and ρ 0.349, compared to baseline MAE 1.539 and ρ 0.105.

    “The best MAE and Spearman’s ρ is obtained using the personalised transfer learning approach with a shared common backbone model and personalised layers with a result of 1.349 and 0.349 compared to the baseline which yields 1.539 and 0.105, respectively.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    DMLI is consistently higher when personalisation is used, while DMUI is consistently lower.

    “DMLI is consistently higher when personalisation is used, while DMUI is consistently lower”
    Quote in the paperSubstantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    Personalisation methods most often improve on both sex fairness metrics (score and bias) for PHQ-2 prediction and forecasting, reducing the bias inherent in the baseline model (which favored females).

    “Notably, personalisation methods most often improve on both metrics,”
    • Also stated in: § Discussion · plain
      “We found that personalisation was able to reduce the bias inherent in the initial baseline model (in our case favoring females).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    All patients obtain improved predictions via personalisation, although not all benefit equally.

    “Similarly, we found that all patients obtain improved predictions via personalisation, although not all benefit equally.”
    • Also stated in: § Individual-level fairness · plain
      “Overall, this leads to a trend where all participants gain by having improved MAE scores.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Subgroup models show the best performance for PHQ-2 prediction, with sex bias almost completely eliminated.

    “with subgroup-models showing the best performance for prediction (where the bias is almost completely eliminated)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Transfer learning with shared layers performs best for PHQ-2 forecasting, where performance remains higher for females but there are no systematic over-/under-predictions.

    “and transfer learning with shared layers performing best for forecasting (where the performance still remains higher for females but there are no systematic over-/under-predictions).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Transfer learning shows better behaviour for DMLI, whereas subgroup models and subject-dependent standardisation work better for DMUI.

    “transfer learning shows better behaviour for DMLI, whereas the other two work better for DMUI.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The baseline model shows good behaviour with respect to GC; it comes first for forecasting with GC 0.161 and second for prediction with GC 0.166 (following subgroup models with 0.148).

    “We note that the baseline model is already showing good behaviour with respect to GC; in fact, it comes first for forecasting with a GC of 0.161 and second for prediction with a GC of 0.166 (following subgroup models with 0.148).”
    Quote in the paperNot a substantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    Transfer learning and subgroup models also improved PHQ-2 prediction over baseline, but less strongly than subject-dependent standardisation.

    “The other two personalisation methods also improved the result, although not quite as strongly as the subject-dependent standardisation method.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    There is sometimes disagreement across methods as to which participants the system works well for; e.g., for forecasting, participant #4 is ranked as the best performing one for all methods except the baseline.

    “Interestingly, there is sometimes disagreement as to where the system works well (e.g., for forecasting, participant #4 is ranked as the best performing one for all methods except the baseline).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Digital interventions are an effective treatment for depression.

    “Digital interventions have been found to be an effective treatment for a broad range of mental disorders (22), including depression (23).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [22] Andersson, 2019, Internet-delivered psychological treatments: from innovation to implementation, World Psychiatry source

      Not checked in the free run.

    • cites [23] Moshe, 2021, Digital interventions for the treatment of depression: a meta-analytic review, Psychol Bull source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Approximately 37% of major depressive disorder patients achieve remission after the first course of evidence-based treatment, and 67% after trying several courses.

    “with approximately 37% achieving remission after the first course of evidence-based treatment (i.e., pharmacotherapy, psychological interventions, or combination therapy), and 67% after trying out several courses of treatment (25).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [25] Rush, 2006, Acute and longer-term outcomes in depressed outpatients requiring one or several treatment steps: a STAR*D report, Am J Psychiatry source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    EMA data has been successfully used to predict the development of symptoms within treatments.

    “However, EMA data has also been successfully used to predict the development of symptoms within treatments [e.g., (42)].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [42] Olthof, 2020, Critical fluctuations as an early-warning signal for sudden gains and losses in patients receiving psychotherapy for mood disorders, Clin Psychol Sci source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    User-dependent models perform better in comparison to user-independent approaches.

    “Li and Sano (51), for example, found that user-dependent models perform better in comparison to user-independent approaches.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [51] Li

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    The results suggest that personalisation using subject-dependent standardisation and transfer learning can improve predictions and forecasts, respectively, of depressive symptoms in participants of a digital depression intervention.

    “Our results suggest that personalisation using subject-dependent standardisation and transfer learning can improve predictions and forecasts, respectively, of depressive symptoms in participants of a digital depression intervention.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Premisecitation

    Fairness requires that model predictions do not show biases for certain protected attributes like race, biological sex, or age.

    “A critical consideration for digital health applications is fairness, which requires that model predictions do not show biases for certain protected attributes like race, biological sex, or age (80).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [80] Corbett-Davies

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The intervention is based on GET.ON Mood Enhancer, a program that has been evaluated in multiple randomised controlled trials.

    “The intervention is based on “GET.ON Mood Enhancer,” a program that has been evaluated in multiple randomised controlled trials (62–64).”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [62] Buntrock, 2016, Effect of a web-based guided self-help intervention for prevention of major depression in adults with subthreshold depression: a randomized clinical trial, JAMA source

      Not checked in the free run.

    • cites [63] Ebert, 2018, Effectiveness of web-and mobile-based treatment of subthreshold depression with adherence-focused guidance: a single-blind randomized controlled trial, Behav Ther source

      Not checked in the free run.

    • cites [64] Reins, 2019, The more I got, the less I need? Efficacy of internet-based guided self-help compared to online psychoeducation for major depressive disorder, J Affect Disord source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The 2- and 9-item Patient Health Questionnaires are psychometrically valid and commonly used measures of depressive symptom severity.

    “The 2- and 9-item Patient Health Questionnaires (PHQ) are two psychometrically valid and commonly used measures of depressive symptom severity (74, 75).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [74] Kroenke, 2003, The patient health questionnaire-2: validity of a two-item depression screener, Med Care source

      Not checked in the free run.

    • cites [75] Löwe, 2004, Measuring depression outcome with a brief self-report instrument: sensitivity to change of the patient health questionnaire (PHQ-9), J Affect Disord source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    There is no clear definition on how to measure fairness for regression tasks, but most approaches try to achieve an equal average expected outcome for different populations.

    “There is no clear definition on how to measure fairness for regression tasks, but most approaches try to achieve an equal average expected outcome for the different populations (55, 81).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [55] Wagner

      Not checked in the free run.

    • cites [81] Fitzsimons

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Telephone-conducted SCID-5 interviews show only slightly inferior reliability compared to face-to-face interviews, supporting their use for research purposes.

    “While the SCID-5 is intended to be delivered face-to-face, interviews conducted via telephone have been found to show only slightly inferior reliability, supporting their use for research purposes (61).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [61] Osório, 2019, Clinical validity, intrarater, test–retest reliability of the structured clinical interview for dsm-5–clinician version (scid-5-cv), Psychiatry Clin Neurosci source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The study procedures were approved by the ethics committee of the Friedrich-Alexander-University Erlangen-Nuremberg (385_20B).

    “The study procedures have been approved by the ethics committee of the Friedrich-Alexander-University Erlangen-Nuremberg (385_20B).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    The MAIKI dataset was collected as part of a prospectively registered feasibility trial (DRKS00024718).

    “The MAIKI dataset was collected as part of a prospectively registered feasibility trial within the MAIKI project (German Clinical Trials Register; DRKS00024718).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The exploratory passive feature list was selected based on technical feasibility and associations with mood symptoms found in previous studies.

    “This exploratory feature list was selected based on technical feasibility, as well as on associations with mood symptoms found in previous studies (38).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [38] Rohani, 2018, Correlations between objective behavioral features collected from mobile and wearable devices and depressive mood symptoms in patients with affective disorders: systematic review, JMIR Mhealth Uhealth source

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning

    The standard individual fairness formulation proposed by Dwork et al. is not applicable in this work because it applies to tasks where each individual is assigned once to a particular class, whereas this study applies models several times to each individual to predict PHQ-2 scores over time.

    “We note that the standard individual fairness formulation proposed by Dwork et al. (57), which purports that similar individuals should receive similar outcomes, is not applicable in our work. This is because Dwork et al. (57) applied their framework to tasks where each individual is assigned once to a particular class (e.g., in recidivism cases); however, we apply our models several times to each”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [57] Dwork

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    A participant's PHQ entry was considered only if the patient provided at least five days of valid data in the week prior to the corresponding PHQ label date.

    “A participant’s PHQ entry is only considered if the patient provides at least five days of valid data in the week prior to the corresponding PHQ label date, taking into account criterion 1 as valid data criterion.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Each study participant needed to provide at least 10 valid entries for PHQ-2 and 5 for PHQ-9.

    “Each study participant needs to provide at least 10 valid entries for PHQ-2 and 5 for PHQ-9 respectively, considering criterion 2.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Missing features for a day with missing rate less than 20% were imputed by calculating the mean value of the missing feature based on all available days in the week prior to the corresponding label of the participant.

    “If features for one day are missing, but the missing rate is less than 20%, we perform statistical imputation by calculating the mean value of the missing feature based on all available days in the week prior to the corresponding label of the participant.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductreasoning

    The study used 3-fold cross validation with subject-dependent data splits, ensuring data from each patient in train, development, and test sets.

    “For evaluating our experiments, we use 3-fold cross validation (CV). To be able to personalise a model to all participants, we need subject-dependent data splits. This means that we need, in all datasets (train/development/test), data from each patient.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Individual fairness arises in several tasks where the target manifests itself differently in distinct human subjects, including fatigue in runners, stress and emotion in voice, and mood prediction from wearable sensors.

    “In addition to group-level fairness, individual fairness arises in several tasks where the target manifests itself differently in distinct (human) subjects. This ranges from the expression of fatigue in runners (54), to stress and emotion in voice (55, 56), and, crucially, mood prediction from wearable sensors (46).”
    Quote in the paperNot a substantial claim0 of 4 sources checked
    • cites [54] Triantafyllopoulos

      Not checked in the free run.

    • cites [55] Wagner

      Not checked in the free run.

    • cites [56] Triantafyllopoulos

      Not checked in the free run.

    • cites [46] Taylor, 2017, Personalized multitask learning for predicting tomorrow’s mood, stress, and health, IEEE Trans Affect Comput source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    The mean-absolute-error was used as evaluation metric because it is more robust to outliers compared to metrics such as mean-squared-error.

    “Furthermore, we use the mean-absolute-error (MAE) as evaluation metric as it is more robust to outliers compared to metrics such as mean-squared-error (MSE).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Group-level fairness analysis focused on biological sex as the only relevant group-level variable available in the study.

    “With regard to fairness at group-level, we focus on biological sex as this is the only relevant, group-level variable available in our study.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Each day with available features of one participant was considered valid only if the data missing rate over all features was less than 20%.

    “Each day with available features of one participant is only considered as valid if the data missing rate over all features is less than 20%.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation

    The affect items were derived from previous personalized sensing studies and a review of affect measurement in previous EMA studies to increase comparability with existing literature.

    “We derived these affect items from previous personalized sensing studies [which used the circumplex model of emotion as basis; (69, 70)] and a review of affect measurement in previous EMA studies (71), to increase the comparability with existing literature.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [69] LiKamWa

      Not checked in the free run.

    • cites [70] Asselbergs, 2016, Mobile phone-based unobtrusive ecological momentary assessment of day-to-day mood: an explorative study, J Med Internet Res source

      Not checked in the free run.

    • cites [71] Scott, 2020, A coordinated analysis of variance in affect in daily life, Assessment source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisenone given

    Spearman's ρ correlation measures how well models are able to predict the correct ordering of instances.

    “In addition, we evaluate our results with Spearman’s ρ correlation, which measures how well models are able to predict the correct ordering of instances—thus enabling medical practitioners to detect the most severe cases.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    The data quality criteria were imposed to maximize the number of participants who contribute information, while trying to minimize the risk of biased results due to large amounts of missing data on a person level.

    “These criteria were imposed to maximize the number of participants who contribute information, while trying to minimize the risk of biased results due to large amounts of missing data on a person level.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The final model state was selected on the basis of validation set performance.

    “The final model state was selected on the basis of validation set performance.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Personalisation allows models to learn patterns specific to each patient's symptomatology, which could explain why performance benefits arise.

    “Personalisation allows models to learn patterns that may be specific to the symptomatology of each patient, which could explain why performance benefits arise.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Taylor et al. reported that personalized Multitask Learning improved prediction accuracies by 11-21% over non-personalized models.

    “Our findings also corroborate the results of Taylor et al. (46), who report that personalized Multitask Learning (shared common layers neural network, multi-kernel learning using support vector machines, and hierarchical Bayesian models with a common Dirichlet process prior) improved prediction accuracies by 11–21% over non-personalized models.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [46] Taylor, 2017, Personalized multitask learning for predicting tomorrow’s mood, stress, and health, IEEE Trans Affect Comput source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Some individuals reported difficulties installing and navigating the companion tracking app.

    “Some individuals, for example, reported difficulties installing and navigating the companion tracking app.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Personalisation approaches increase the diversity of performance within individual participants, seemingly increasing inequality.

    “This shows that our proposed personalisation approaches increase the diversity of performance within individual participants, thus seemingly increasing inequality.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Missing values leading to subset inclusion is a common finding in real-world data studies, particularly with mobile sensor features.

    “This is a common finding in studies based on real-world data, particularly if mobile sensor features are included [see, e.g., (34, 35)].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [34] Saeb, 2015, Mobile phone sensor correlates of depressive symptom severity in daily-life behavior: an exploratory study, J Med Internet Res source

      Not checked in the free run.

    • cites [35] Pratap, 2019, The accuracy of passive phone sensors in predicting daily mood, Depress Anxiety source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    The sex fairness metrics show that males are predicted more wrongly with a bias towards negatives, causing many high PHQ-2 cases to be mispredicted as low.

    “Collectively, these two metrics show that males are predicted more wrongly with a bias towards negatives—causing a lot of high PHQ-2 cases to be mispredicted as having low PHQ-2 scores.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The DMLI/DMUI pattern shows that 'winners' further diverge from the median while 'losers' come closer to it.

    “this shows that “winners” further diverge from the median while “losers” come closer to it.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given

    A wide variance of PHQ values must be collected from each participant during training; otherwise, the personalised model may tend to overfit.

    “It is also important to note that a wide variance of PHQ values must be collected from each participant during this time—otherwise, the personalised model may tend to overfit.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Findings provide preliminary evidence that model personalisation, based on EMA ratings and supported by passive mobile data, can improve daily forecasts of depressive symptom severity in real-world patient populations, as well as their fairness.

    “In sum, our findings provide preliminary evidence that model personalisation, based on EMA ratings and supported by passive mobile data, can be used to improve daily forecasts of depressive symptom severity in real-world patient populations, as well as their fairness.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The subset of patients with sufficient active EMA and passive mobile sensor data is representative of the broader patient population with elevated depressive symptoms.

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    EMA ratings may serve a twofold objective: providing therapeutic feedback to patients and allowing building a personalised forecasting model of the individual's symptomatology.

    “and EMA ratings may therefore serve a twofold objective: providing therapeutic feedback to patients, while at the same time allowing to build a personalised forecasting model of the individual’s symptomatology.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Structured self-monitoring is a common feature in cognitive-behavioural treatments for depression.

    “Structured self-monitoring is a common feature in cognitive-behavioural treatments for depression (87)”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [87] Cohen, 2013, Using self-monitoring: implementation of collaborative empiricism in cognitive-behavioral therapy, Cogn Behav Pract source

      Not checked in the free run.

52 of 52

  • The paper relies on this without support.

    Predictive models, especially general ones, often work better in one subgroup of the population than in others.

    “Predictive models, especially general ones, often work better in one subgroup of the population than in others.”

  • The paper relies on this without support.

    There is no widely-accepted metric to quantify individual fairness.

    “However, there is no widely-accepted metric to quantify individual fairness.”

  • The paper relies on this without support.

    Current digital interventions typically follow a one size fits all approach that is very limited in its capability to react adaptively to patients’ individual trajectory.

    “Current digital interventions typically follow a “one size fits all” approach that is very limited in its capability to react adaptively to patients’ individual trajectory.”

  • The paper relies on this without support.

    Digital interventions can be provided to everyone with Internet access and are highly scalable.

    “Digital interventions can be provided to everyone with Internet access and are highly scalable.”

  • The paper relies on this without support.

    Model fairness across patient subgroups is frequently overlooked.

    “Model fairness across patient subgroups is also frequently overlooked.”

  • The paper relies on this without support.

    It is still largely unclear how patients’ individual symptoms evolve dynamically during such treatments.

    “but it is still largely unclear how patients’ individual symptoms evolve dynamically during such treatments.”

  • The paper relies on this without support.

    The subset of patients with sufficient active EMA and passive mobile sensor data is representative of the broader patient population with elevated depressive symptoms.

    “First, due to missing values, only a subset of patients in the MAIKI dataset could be included in our experiments.”
7 of 7

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?

  • Different individuals may exhibit varying symptoms or responses to the same treatment, making a case for improved personalization.

    “Additionally, different individuals may exhibit varying symptoms or responses to the same treatment. This requires meticulous attention for systems designed around individual needs, making a case for improved personalization [189].”

    From Human-centred artificial intelligence for mobile health sensing: challenges and opportunities (2023) · cites this paper as [189]

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.

  • Sex fairness score sign is interpreted backwards

    Major

    The defined score is female MAE minus male MAE. A positive score therefore indicates higher MAE for females, not males; a negative score indicates higher MAE for males, not females.

  • Baseline sex fairness bias is described as negative but is positive

    Major

    The text calls the baseline bias negative for both tasks, whereas Table 4 reports positive baseline biases of 0.576 for prediction and 0.630 for forecasting.

  • Claim that the Gini coefficient decreases conflicts with Table 5

    Minor

    Table 5 shows that the Gini coefficient increases relative to baseline for transfer learning and subject-dependent standardisation in both tasks, and for subgroup forecasting. Only subgroup prediction has a lower coefficient.

  • Transfer learning does not have the best forecasting DMLI

    Minor

    The closing comparison says transfer learning performs better on DMLI, but Table 5 reports a forecasting DMLI of 3.377 for subject-dependent standardisation versus 2.664 for transfer learning. The paper states that higher DMLI is better.

4 of 4

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. First, 2016, User’s guide for the SCID-5-CV structured clinical interview for DSM-5®disorders: clinical version

Cited by

Sorted by citations

Related

Authors
Kathan, A., Harrer, M., Küster, L., Triantafyllopoulos, A., He, X., Milling, M., Gerczuk, M., Yan, T., Rajamani, S. T., Heber, E., Grossmann, I., Ebert, D. D., & Schuller, B. W.
Affiliations
  • University of Augsburg (DE)
  • Friedrich-Alexander-Universität Erlangen-Nürnberg (DE)
  • Technical University of Munich (DE)
Corresponding author
Kathan, Alexander
ORCID
Venue
Frontiers in Digital Health
Publisher
Frontiers Media SA
Published
2022-11-18
Volume
4
Pages
964582-964582
ISSN
2673-253X (ISSN-L)
DOI
10.3389/fdgth.2022.964582
PMC
PMC9715619
OpenAlex
W4309271221
Semantic Scholar
24a7c772b5d4c595a010c161def019758b423672
Cited by
47
References
79
Citations by year
2026: 14 · 2025: 14 · 2024: 11 · 2023: 8
FWCI
7.59
Type
journal article
Language
en
Open access
Licence
Locations
  • Frontiers in Digital Health (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) (submittedVersion, other-oa)
  • mediaTUM (Technical University of Munich) (submittedVersion, other-oa)
  • OPUS Repository (Kooperativer Bibliotheksverbund Berlin-Brandenburg) (submittedVersion, cc-by)
  • OPUS (Augsburg University) (submittedVersion, cc-by)
Funders
Bundesministerium für Bildung und Forschung
Topics
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Mental Health Research Topics
  • Social Sciences › Psychology › Applied Psychology › Digital Mental Health Interventions
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Emotion and Mood Recognition
Keywords
depression, mobile sensor data, ecological momentary assessment, digital health interventions, transfer learning, model fairness, PHQ-2, day-ahead forecasting

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 2026-10-08.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-08.

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