CoChat
Sign Up
Cited by 109Machine Learning in HealthcareExplainable Artificial Intelligence (XAI)Artificial Intelligence in Healthcare

Using explainable machine learning to characterise data drift and detect emergent health risks for emergency department admissions during COVID-19

Open access
Duckworth, Christopher, Chmiel, Francis P., Burns, Dan K., Zlatev, Zlatko D., White, Neil M., Daniels, Thomas W. V., Kiuber, Michael, Boniface, Michael J.·2021·Scientific Reports
DOI 10.1038/s41598-021-02481-yPMC PMC8626460

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

A key task of emergency departments is to promptly identify patients who require hospital admission. Early identification ensures patient safety and aids organisational planning. Supervised machine learning algorithms can use data describing historical episodes to make ahead-of-time predictions of clinical outcomes. Despite this, clinical settings are dynamic environments and the underlying data distributions characterising episodes can change with time (data drift), and so can the relationship between episode characteristics and associated clinical outcomes (concept drift). Practically this means deployed algorithms must be monitored to ensure their safety. We demonstrate how explainable machine learning can be used to monitor data drift, using the COVID-19 pandemic as a severe example. We present a machine learning classifier trained using (pre-COVID-19) data, to identify patients at high risk of admission during an emergency department attendance. We then evaluate our model’s performance on attendances occurring pre-pandemic (AUROC of 0.856 with 95%CI [0.852, 0.859]) and during the COVID-19 pandemic (AUROC of 0.826 with 95%CI [0.814, 0.837]). We demonstrate two benefits of explainable machine learning (SHAP) for models deployed in healthcare settings: (1) By tracking the variation in a feature’s SHAP value relative to its global importance, a complimentary measure of data drift is found which highlights the need to retrain a predictive model. (2) By observing the relative changes in feature importance emergent health risks can be identified.

AI snapshot

Written from the paper’s full text.

TL;DR

The authors trained an XGBoost model on pre-COVID-19 emergency department attendances to predict hospital admission and used SHAP explanations to monitor data drift during the COVID-19 pandemic. Model discrimination fell from AUROC 0.856 pre-pandemic to 0.826 during the pandemic, while SHAP feature tracking highlighted changes in patient behaviour, operational procedure, and possible emergent health risks.

Key findings

  • The admission model achieved AUROC 0.856 (95%CI [0.852, 0.859]) on the pre-COVID-19 test set and AUROC 0.826 (95%CI [0.814, 0.837]) during the COVID-19 pandemic.
  • During COVID-19, average precision increased misleadingly because more attendees were admitted; at fixed recall 0.75, performance degradation appeared mainly as a higher false-alarm rate among patients who were actually discharged.
  • Normalised SHAP values showed increased relative importance during COVID-19 for respiration rate, triage discriminator, and arrival mode, and decreased relative importance for triage complaint, attendance complaint, and comorbidity count.
  • SHAP tracking identified possible emergent health risks: respiration rate importance peaked in December 2019 and March 2020, and triage discriminator categories such as pleuritic pain and low SpO2 became more prevalent, likely encoding symptomatic COVID-19.

Methods

  • Retrospective study of 82,402 adult ED attendances at Southampton General Hospital from 1 April 2019 to 30 April 2020, split temporally into training (40,309), pre-COVID-19 test (37,367), and COVID-19 test (4,726) cohorts.
  • An XGBoost classifier predicted admission versus discharge at triage using encoded demographics, chief complaints, attendance characteristics, triage vital signs, and one-hot encoded prior discharge conditions; hyperparameters were selected by fivefold cross-validation with Optuna.
  • Performance was evaluated with AUROC and precision-recall curves; SHAP TreeExplainer values were used for local and global feature attribution, with absolute normalised SHAP values tracked weekly to compare relative feature importance over time.

Limitations

  • The paper states that prior patient information came only from previous discharge summaries at UHS, so patient histories were available for only 55% of recorded attendances, and that the work cannot formally detect concept drift, only data drift.
  • Evident from the text: this is a single-centre retrospective study, the model had no access to free-text fields, and the COVID-19 test cohort was relatively small and limited to the first pandemic wave.
  • Evident from the text: a model trained pre-COVID-19 would treat newly added categories such as “COVID-19” as missing information, which could affect optimal use of recorded features.

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

Integrity signals

Registry

Crossref · 10.1038/s41598-021-02481-y

Cites

Statistics

Nothing to recompute.

Claims the paper makes

0 of 26 cited sources checked · 0 supported · 0 flagged

  • Findingreasoning

    The onset of the COVID-19 pandemic led to a sudden uptick in the fraction of ED attendances with recorded vital signs at UHS.

    “Specifically the onset of the COVID-19 pandemic led to a drop in ED attendances which led to sudden uptick in admission rate, average age of patients attending, and the fraction of attendances with recorded vital signs (see Fig. 2).”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The COVID-19 test cohort is representative of ED attendances during the COVID-19 pandemic at UHS.
  • Findingown data

    There was a significant decrease in patients attending through walk-ins during the first wave of the COVID-19 pandemic.

    “As seen in Fig. 2e we found a significant decrease in patients attending through ‘walk-ins’, which leads to increased SHAP values for the feature because other modes of arrivals (e.g., ambulance arrivals) are generally used by patients who are more severely ill and are therefore more likely to be admitted.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The COVID-19 test cohort is representative of ED attendances during the COVID-19 pandemic at UHS.
  • Findingreasoning

    The model's average precision under the precision-recall curve increased during the COVID-19 pandemic.

    “Again, we see a statistically significant drop in the AUROC during the COVID-19 pandemic, but, the model’s average precision under the precision-recall curve actually increases during this time.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The historical discharge summaries and electronic health record fields used as input features accurately and completely capture each patient's relevant medical history.
    • Rests on an assumption: The recorded admission or discharge outcome accurately reflects the patient's true clinical outcome.
    • Rests on an assumption: The model's predicted admission risk is a valid proxy for the clinical need for hospital admission.
    • Rests on an assumption: The COVID-19 test cohort is representative of ED attendances during the COVID-19 pandemic at UHS.
  • Findingreasoning

    The onset of the COVID-19 pandemic led to a sudden uptick in the average age of patients attending UHS ED.

    “Specifically the onset of the COVID-19 pandemic led to a drop in ED attendances which led to sudden uptick in admission rate, average age of patients attending, and the fraction of attendances with recorded vital signs (see Fig. 2).”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The COVID-19 test cohort is representative of ED attendances during the COVID-19 pandemic at UHS.
  • Findingreasoning

    The onset of the COVID-19 pandemic led to a sudden uptick in the admission rate at UHS.

    “Specifically the onset of the COVID-19 pandemic led to a drop in ED attendances which led to sudden uptick in admission rate, average age of patients attending, and the fraction of attendances with recorded vital signs (see Fig. 2).”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The COVID-19 test cohort is representative of ED attendances during the COVID-19 pandemic at UHS.
  • May combine two claims. 25
  • The quote only partly states this claim. 1
  • Classified as probably not a substantial claim. 15

  • May combine two claims.

    Findingown data

    The drop in ED attendances during the onset of COVID-19 was driven by patients no longer walking in and by a reduction in attendances attributed to wounds or related traumas.

    “The drop of attendances appear to be driven by patients no longer ‘walking-in’ into the ED and because of a reduction in the number of attendances attributed to ‘wounds’ or related traumas (i.e. head injuries or limb problems).”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The COVID-19 test cohort is representative of ED attendances during the COVID-19 pandemic at UHS.

  • May combine two claims.

    Findingown data

    Respiration rate and triage discriminator decreased in average normalised SHAP importance post April 2020, while arrival mode remained consistently more important than pre-COVID-19.

    “While both respiration rate and triage discriminator decrease in their average importance post April 2020, arrival mode remains consistently more important than pre-COVID-19.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The historical discharge summaries and electronic health record fields used as input features accurately and completely capture each patient's relevant medical history.
    • Rests on an assumption: The recorded admission or discharge outcome accurately reflects the patient's true clinical outcome.
    • Rests on an assumption: The model's predicted admission risk is a valid proxy for the clinical need for hospital admission.
    • Rests on an assumption: The COVID-19 test cohort is representative of ED attendances during the COVID-19 pandemic at UHS.

  • May combine two claims.

    Findingreasoning

    During COVID-19, the importance of 'triage discriminator' increased and the importance of 'attendance (and triage) complaint' decreased for predicting admission risk.

    “Another significant shift is the increased importance (during COVID-19) of ‘triage discriminator’ and the decrease in ‘attendance (and triage) complaint’.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For triage discriminator, specific categories such as pleuritic pain and (very) Low SpO2 became far more prevalent during COVID-19, whereas uninformative categories (e.g. recent problem, recent mild pain) were far more common pre-COVID-19.

    “Conversely for triage discriminator (right panel), specific categories such as pleuritic pain and (very) Low SpO2 directly related to viral infections become far more prevalent, whereas uninformative categories (e.g. recent problem, recent mild pain) were far more common pre COVID-19.”
    Quote in the paperSubstantial claim

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

    Findingown data

    The model's performance during COVID-19 was AUROC 0.826 (95% CI [0.814, 0.837]), a small degradation from pre-COVID-19.

    “We demonstrate that this significant circumstantial change during the first wave of the pandemic is correlated with a small degradation of model performance (AUROC = 0.83; bottom left in Fig. 4).”
    • Also stated in: § Conclusion · plain
      “Using the COVID-19 pandemic as an exemplar of data drift, we evaluated our model’s performance during severe operational change with our classifier obtaining an AUROC of 0.826 with 95%CI [0.814, 0.837].”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For attendance and triage complaints, generic and uninformative categories (e.g. unwell adult) were significantly more prevalent during COVID-19, while specific problems such as head injuries and wounds were far less prevalent.

    “Generic and uninformative categories are significantly more prevalent during COVID-19 for the complaints (e.g. unwell adult), whereas specific problems such as head injuries and wounds are far less prevalent.”
    Quote in the paperSubstantial claim

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

    Interpretationcitation

    The observed changes in ED attendance behaviour at UHS are reflective of national attendance behaviour changes in the UK through the onset of the pandemic.

    “This is reflective of national attendance behaviour changes in the UK through the onset of the pandemic8.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [8] Kelly, E. & Firth, Z. How is covid-19 changing the use of emergency care? https://www.health.org.uk/news-and-comment/charts-and-infographics/how-is-covid-19-changing-the-use-of-emergency-care (2020). Internet resource.

      Not checked in the free run.

    • Rests on an assumption: The COVID-19 test cohort is representative of ED attendances during the COVID-19 pandemic at UHS.

  • May combine two claims.

    Premisecitation

    SHAP offers consistency between local explanations, enabling aggregation for global interpretation.

    “SHAP offers consistency between local explanations enabling aggregation for global interpretation.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [18] Lundberg, S. M. et al. Explainable AI for trees: From local explanations to global understanding. arXiv preprint arXiv:1905.04610 (2019).

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    On-the-fly model evaluation, such as AUROC, along with normalised SHAP values would indicate both a degradation in performance and attribution of what is changing, such as features associated with clinical procedure or patient attendance behaviour.

    “On-the-fly model evaluation (i.e. through a performance metric such as AUROC), along with normalised SHAP values would not only indicate a degradation in performance, but also, attribution of what is changing (e.g. features associated with clinical procedure or patient attendance behaviour).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The historical discharge summaries and electronic health record fields used as input features accurately and completely capture each patient's relevant medical history.
    • Rests on an assumption: The recorded admission or discharge outcome accurately reflects the patient's true clinical outcome.
    • Rests on an assumption: The model's predicted admission risk is a valid proxy for the clinical need for hospital admission.
    • Rests on an assumption: The COVID-19 test cohort is representative of ED attendances during the COVID-19 pandemic at UHS.

  • May combine two claims.

    Interpretationreasoning

    Respiration rate and triage discriminator are likely reflective of dynamically changing ED operations and highlight procedural change or record-keeping as the COVID-19 pandemic progressed.

    “Notably, respiration rate and triage discriminator are likely reflective of dynamically changing ED operations, and, highlight procedural change (or record-keeping) as the COVID-19 pandemic progressed.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The historical discharge summaries and electronic health record fields used as input features accurately and completely capture each patient's relevant medical history.
    • Rests on an assumption: The recorded admission or discharge outcome accurately reflects the patient's true clinical outcome.
    • Rests on an assumption: The model's predicted admission risk is a valid proxy for the clinical need for hospital admission.
    • Rests on an assumption: The COVID-19 test cohort is representative of ED attendances during the COVID-19 pandemic at UHS.

  • May combine two claims.

    Premisecitation

    SHAP offers theoretical guarantees of trustworthiness, such as local accuracy, missingness, and consistency, which adhere to strict medical governance.

    “SHAP offers theoretical guarantees of trustworthiness (e.g. local accuracy, missingness, and consistency) which adhere to strict medical governance18”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [18] Lundberg, S. M. et al. Explainable AI for trees: From local explanations to global understanding. arXiv preprint arXiv:1905.04610 (2019).

      Not checked in the free run.

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

    Premisecitation

    Gradient boosted trees combine an ensemble of weak learners to decrease bias while preserving or lowering variance in prediction error, making them more desirable over other tree-based algorithms.

    “Gradient boosted trees are examples of boosting algorithms which aim to combine an ensemble of weak learners in order to decrease bias whilst preserving or lowering variance in the prediction error14; this makes them more desirable over other tree-based algorithms.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [14] Breiman, L. Bias, Variance, and Arcing Classifiers (Tech, Rep, 1996).

      Not checked in the free run.

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

    Premisecitation

    Models that predict patient outcomes, such as discharge or admission, seek to improve logistics in EDs and organise downstream resource allocation.

    “Models which predict patient outcomes (e.g. whether a patient is likely to be discharged or admitted to hospital) seek to improve logistics in EDs and organise downstream resource allocation9,10.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [9] Arya, R. et al. Decreasing length of stay in the emergency department with a split emergency severity index 3 patient flow model. Acad. Emerg. Med. 20, 1171–1179. https://doi.org/10.1111/acem.12249 (2013). source

      Not checked in the free run.

    • cites [10] Haimovich, J. S. et al. Discovery of temporal and disease association patterns in condition-specific hospital utilization rates. PLoS ONE 12, 1–15. https://doi.org/10.1371/journal.pone.0172049 (2017). source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisereasoning

    For machine learning models used as decision support tools, data drift can invalidate a model because the drifted feature distribution may predominantly lie in a region where the model poorly performs.

    “For machine learning models, used as decision support tools, data drift can invalidate a model, as the drifted feature distribution may predominantly lie in a region where the model poorly performs.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    Model performance was evaluated independently on both test sets, pre-COVID-19 and during COVID-19, using AUROC and the precision-recall curve.

    “Model performance was evaluated independently by both test sets (pre and during COVID-19) using the Area Under the Receiver Operating Curve (AUROC) and with the precision recall curve.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    Shapley values are consistent, meaning the values add up to the actual prediction of the model, so they can be used to check the global importance of a feature.

    “In addition, Shapley values are consistent (the values add up to the actual prediction of the model) meaning they can also be used to check the global importance of a feature.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [16] Lundberg, S. M. et al. Explainable machine learning predictions to help anesthesiologists prevent hypoxemia during surgery. bioRxiv https://doi.org/10.1101/206540 (2017). https://www.biorxiv.org/content/early/2017/10/21/206540.full.pdf. source

      Not checked in the free run.

    • cites [17] Lundberg, S. M. & Lee, S.-I. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, Vol. 30 (eds Guyon, I. et al.) (Curran Associates, Inc., 2017).

      Not checked in the free run.

    • cites [18] Lundberg, S. M. et al. Explainable AI for trees: From local explanations to global understanding. arXiv preprint arXiv:1905.04610 (2019).

      Not checked in the free run.

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

    Interpretationreasoning

    Predicting patient outcomes at the point of triage enables automation within the ED triaging system, ahead-of-time resource prediction, and downstream logistical planning.

    “Predicting patient outcomes at this point in a patient’s care trajectory enables automation within the triaging system of the ED itself, ahead-of-time resource prediction, and downstream logistical planning.”
    Quote in the paperNot a substantial claim

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

    Study design and conductcitation

    Model hyperparameters were selected using fivefold cross-validation of the complete training set using a Tree-structured Parzen Estimator implemented with Optuna.

    “Model hyperparameters were selected using fivefold cross-validation of the complete training set using a sampler (Tree-structured Parzen Estimator) implemented with the Optuna library15.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [15] Akiba, T., Sano, S., Yanase, T., Ohta, T. & Koyama, M. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2019). source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    Shapley values are locally accurate, meaning they can explain a prediction for a single attendance.

    “One particular strength of Shapley values is that they are locally accurate, i.e, they can explain a prediction for a single attendance.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [16] Lundberg, S. M. et al. Explainable machine learning predictions to help anesthesiologists prevent hypoxemia during surgery. bioRxiv https://doi.org/10.1101/206540 (2017). https://www.biorxiv.org/content/early/2017/10/21/206540.full.pdf. source

      Not checked in the free run.

    • cites [17] Lundberg, S. M. & Lee, S.-I. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, Vol. 30 (eds Guyon, I. et al.) (Curran Associates, Inc., 2017).

      Not checked in the free run.

    • cites [18] Lundberg, S. M. et al. Explainable AI for trees: From local explanations to global understanding. arXiv preprint arXiv:1905.04610 (2019).

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    The increase in average precision during COVID-19 was not due to increased discriminative ability but to a change in the target variable distribution, with the average patient attending during COVID-19 being more likely to be admitted.

    “This boost in precision, however, is not a reflection of increased discriminative ability between admission or discharge for the model, but actually the result of a change in the target variable distribution: the average patient attending in the COVID-19 pandemic is far more likely to be admitted (Fig. 2b) during this time period.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The historical discharge summaries and electronic health record fields used as input features accurately and completely capture each patient's relevant medical history.
    • Rests on an assumption: The recorded admission or discharge outcome accurately reflects the patient's true clinical outcome.
    • Rests on an assumption: The model's predicted admission risk is a valid proxy for the clinical need for hospital admission.
    • Rests on an assumption: The COVID-19 test cohort is representative of ED attendances during the COVID-19 pandemic at UHS.

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

    Premisenone given

    The area under the precision-recall curve is an inherently biased metric because it depends on the class distribution.

    “We note, that due to there being a significant shift in admission rate between test samples (Fig. 2b), and the area under the precision recall curve is an inherently biased metric (i.e. depends on the class distribution), and, we further discuss its role in data drift interpretation in the results section.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    TreeExplainer, as implemented in the SHAP library, efficiently calculates Shapley values for attributing the fractional contribution of an input feature to admission risk.

    “To attribute the fractional contribution of a given input feature to admission risk, we made use of the TreeExplainer algorithm as implemented in the SHAP (SHapley Additive exPlanations) library16–18.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [16] Lundberg, S. M. et al. Explainable machine learning predictions to help anesthesiologists prevent hypoxemia during surgery. bioRxiv https://doi.org/10.1101/206540 (2017). https://www.biorxiv.org/content/early/2017/10/21/206540.full.pdf. source

      Not checked in the free run.

    • cites [17] Lundberg, S. M. & Lee, S.-I. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, Vol. 30 (eds Guyon, I. et al.) (Curran Associates, Inc., 2017).

      Not checked in the free run.

    • cites [18] Lundberg, S. M. et al. Explainable AI for trees: From local explanations to global understanding. arXiv preprint arXiv:1905.04610 (2019).

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationreasoning

    Identifying patients at highest risk of admission provides a sanity check for triage and helps ensure high acuity patients have timely medical intervention.

    “Identifying patients at highest risk of admission, provides a sanity check for triage and helps ensure high acuity patients have timely medical intervention.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The decrease in walk-in attendances leads to increased SHAP values for arrival mode because other modes of arrival, such as ambulance, are generally used by patients who are more severely ill and are therefore more likely to be admitted.

    “As seen in Fig. 2e we found a significant decrease in patients attending through ‘walk-ins’, which leads to increased SHAP values for the feature because other modes of arrivals (e.g., ambulance arrivals) are generally used by patients who are more severely ill and are therefore more likely to be admitted.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The historical discharge summaries and electronic health record fields used as input features accurately and completely capture each patient's relevant medical history.
    • Rests on an assumption: The recorded admission or discharge outcome accurately reflects the patient's true clinical outcome.
    • Rests on an assumption: The model's predicted admission risk is a valid proxy for the clinical need for hospital admission.
    • Rests on an assumption: The COVID-19 test cohort is representative of ED attendances during the COVID-19 pandemic at UHS.

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

    Premisenone given

    The XGBoost implementation naturally deals with continuous, binary or discrete, and missing data consistently, all of which are represented in the dataset.

    “Additionally, the XGBoost implementation naturally deals with continuous, binary/discrete, and, missing data consistently; all of which are represented in our dataset.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Using explainable machine learning, we are able to determine the importance of input features per attendance, enabling identification of data drift.

    “Using explainable machine learning, we are able to determine the importance of input features, per attendance, enabling identification of data drift.”
    Quote in the paperSubstantial claim

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

    Interpretationreasoning

    Feature importance can be checked periodically by averaging over a fixed time period.

    “Feature importance can therefore be checked periodically by averaging over a fixed time period.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Admission risk appears to be largely dictated by patient frailty and attendance complaint, and therefore admission models implemented early in the ED visit pathway are still highly predictive.

    “Admission risk appears to be largely dictated by patient frailty and attendance complaint, and, hence admission models implemented early in the ED visit pathway are still highly predictive.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    SHAP directly quantifies how impactful data drift has been per covariate for the model, rather than only assessing the magnitude of change, and hence provides more insight into whether a model requires retraining.

    “SHAP directly quantifies how impactful the data drift has been per covariate for the model, rather than only assessing the magnitude of change, and hence, provides more insight into whether a model requires retraining.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    For a model trained pre-COVID-19, any patient designated under 'COVID-19' would be treated as missing information for the category, which the model would not know how to optimally use.

    “For a model trained pre-COVID-19, any patient designated under ‘COVID-19’ would be treated as missing information for the category, which the model would not know how to optimally use.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Arrival mode likely became more important because the average severity of attendances increased due to a significant drop in patients attending through walk-ins.

    “Due to the significant drop of patients attending through ‘walk-ins’ (Fig. 2; typically associated with less frail attendances), arrival mode likely became more important as the average severity of attendances increased.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Patients attending through walk-ins are typically less frail.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Tracking how a model uses information over time can enable a complementary assessment of whether a model needs retraining.

    “However, these act externally to the classifier, and tracking how a model uses the information provided to it over time, could enable a complimentary assessment of whether a model needs retraining.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductown data

    The study only possesses prior patient information from previous discharge summaries at UHS, and therefore only has patient histories for 55% of recorded attendances.

    “One limitation of this study is that we only possess prior patient information from previous discharge summaries at UHS, and, hence only have patient histories for 55% of our recorded attendances.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Triage discriminator and vital signs are reflective of procedural change.

    “Conversely triage discriminator, and, vital signs are reflective of procedural change.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationcitation

    The model's performance (AUROC 0.856) is state-of-the-art and comparable with recent admission models (AUROC 0.825-0.877).

    “This represents state-of-the-art performance comparable with recent admission models (e.g. AUROC 0.825–0.87722–25).”
    Quote in the paperNot a substantial claim0 of 4 sources checked
    • cites [22] Sun, Y., Heng, B. H., Tay, S. Y. & Seow, E. Predicting hospital admissions at emergency department triage using routine administrative data. Acad. Emerg. Med. 18, 844–850. https://doi.org/10.1111/j.1553-2712.2011.01125.x (2011). source

      Not checked in the free run.

    • cites [23] Cameron, A., Rodgers, K., Ireland, A., Jamdar, R. & McKay, G. A. A simple tool to predict admission at the time of triage. Emerg. Med. J. 32, 174–179 (2015). source

      Not checked in the free run.

    • cites [24] Hong, W. S., Haimovich, A. D. & Taylor, R. A. Predicting hospital admission at emergency department triage using machine learning. PLoS ONE 13, 1–13. https://doi.org/10.1371/journal.pone.0201016 (2018). source

      Not checked in the free run.

    • cites [25] Parker, C. A. et al. Predicting hospital admission at the emergency department triage: A novel prediction model. Am. J. Emerg. Med. 37, 1498–1504. https://doi.org/10.1016/j.ajem.2018.10.060 (2019). source

      Not checked in the free run.

35 of 35

  • The paper relies on this without support.

    The historical discharge summaries and electronic health record fields used as input features accurately and completely capture each patient's relevant medical history.

    “We provided the classifier with a view of a patient’s medical history by making use of the historical discharge summaries associated with the patient, both from the ED and from the patients electronic health record.”

  • The paper relies on this without support.

    The recorded admission or discharge outcome accurately reflects the patient's true clinical outcome.

    “We considered only attendances resulting in admission or discharge (i.e. removing planned visits or referrals from other EDs), of which there are 82,402.”

  • The paper relies on this without support.

    The model's predicted admission risk is a valid proxy for the clinical need for hospital admission.

    “The aim of our admissions model was to predict patient outcomes (admitted or discharged) at the point of triage during an ED attendance.”

  • The paper relies on this without support.

    The COVID-19 test cohort is representative of ED attendances during the COVID-19 pandemic at UHS.

    “We split our data temporally into training (40,309 attendances), pre COVID-19 test (37,367 attendances) and COVID-19 test (4726 attendances) cohorts, as detailed in Fig. 1.”

  • The paper relies on this without support.

    A COVID-19 category was not present in the ED's record keeping system in the initial stages of the COVID-19 pandemic.

    “During the COVID-19 pandemic additional complaint categories (e.g. ‘COVID-19’) were eventually added to hospital record keeping systems.”
    “Attendance/triage complaint more frequently contained uninformative categories (i.e. ‘unwell adult’) due to a lack of a COVID-19 category (not added to the ED’s record keeping system in the initial stages of the COVID-19 pandemic).”

  • The paper relies on this without support.

    Patients attending through walk-ins are typically less frail.

    “Due to the significant drop of patients attending through ‘walk-ins’ (Fig. 2; typically associated with less frail attendances), arrival mode likely became more important as the average severity of attendances increased.”

  • The paper relies on this without support.

    Within UHS, a combination of vital signs were used to assess COVID-19 risk.

    “Within UHS, a combination of vital signs were used to assess COVID-19 risk.”
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?

  • Data drift can cause models' performance to deteriorate or behave in unexpected ways.

    “Data drift, defined as the systematic shift in the underlying distribution of input features 11 , can cause models' performance to deteriorate [11][12][13] or behave in unexpected ways, which can pose a threat to the safety of patients.”

    From Empirical data drift detection experiments on real-world medical imaging data (2024) · cites this paper as [11]

The citation lists haven’t been fetched yet

References 37 · Cited by 109

Authors
Duckworth, C., Chmiel, F. P., Burns, D. K., Zlatev, Z. D., White, N. M., Daniels, T. W. V., Kiuber, M., & Boniface, M. J.
Affiliations
  • University of Southampton (GB)
  • Southampton General Hospital (GB)
  • University Hospital Southampton NHS Foundation Trust (GB)
Corresponding author
Duckworth, Christopher
ORCID
Venue
Scientific Reports
Publisher
Springer Science and Business Media LLC
Published
2021-11-26
Volume
11
Issue
1
Pages
23017-23017
ISSN
2045-2322 (ISSN-L)
DOI
10.1038/s41598-021-02481-y
PMC
PMC8626460
OpenAlex
W3217489051
Semantic Scholar
e90d6e7b5971a2b1645e05ae0dba6188c92aab28
Cited by
109
References
37
Citations by year
2026: 18 · 2025: 28 · 2024: 23 · 2023: 27 · 2022: 12 · 2021: 1
FWCI
5.65
Type
journal article
Language
en
Open access
Licence
Locations
  • Scientific Reports (publishedVersion, cc-by)
  • medRxiv (acceptedVersion)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • ePrints Soton (University of Southampton) (acceptedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Alan Turing Institute
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
Keywords
data drift, concept drift, explainable machine learning, SHAP, emergency department admission, COVID-19 pandemic, model monitoring, feature importance, hospital admission prediction, model retraining
MeSH terms
Machine Learning, COVID-19, SARS-CoV-2, Algorithms, Emergency Service, Hospital, Female, Hospitalization, Humans, Male, Middle Aged, Patient Admission, Pandemics

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