Concurrent prediction of in-hospital mortality and length of stay using single-task, multiclass and multitask machine learning

Figures

• 

  Figure 1. Clinician predictions of discharge. (a, b) Time of the day of clinician prediction. (c, d) Days from admission to when clinicians made predictions (prediction day). (e, f) Days from prediction day to actual discharge day. (g, h) Days from prediction day to predicted discharge day. (i, j) Days between clinician prediction and actual discharge. (a), (c), (e), (g) and (i) are elective admissions, while (b), (d), (f), (h) and (j) are emergency admissions.

• 

  Figure 2. Comparison of model performance. Extreme gradient boosting (XGB), logistic regression (LR), multilayer perceptron (MLP), TabNet and multitask learning model were built for predicting mortality and future length of stay (LoS), and the multiclass XGB model was built for predicting combined mortality and LoS. *Macro-averaged performance for mortality classes (died in hospital) or LoS classes (Q1–Q4) from the multiclass XGB model predicting combined mortality and LoS. (a) Area under the receiver-operating curve (AUROC) and area under the precision-recall curve (AUPRC) for elective admissions. (b) AUROC and AUPRC for emergency admissions. Means with 95% CIs are calculated using bootstrap. In-hospital mortality occurred following 2982 (1.6%) elective and 79 646 (8.3%) emergency patient-days. Across all patient-days studied, the median future LoS was 3.2 (IQR 1.1–10.0) days and 5.9 (1.9–14.2) days for elective and emergency admissions, respectively (ie, longer than the overall LoS as longer admissions contributed proportionally more to the dataset than shorter admissions). In elective admissions, patient-days were distributed as 0–1 day (Q1, 20.1%), 2–3 days (Q2, 25.9%), 4–10 days (Q3,…

• 

  Figure 3. Changes in model performance and predicted probabilities by time in the test data using extreme gradient boosting models. (a) Changes in AUROC and AUPRC by time since admission for predicting mortality. (b) Changes in AUROC and AUPRC by time to discharge for predicting mortality. (c) Changes in predicted probability of mortality by time to discharge in patients who discharged alive and died in hospital (individual predicted trajectories are shown in online supplemental figure S4). (d) Changes in AUROC and AUPRC by time since admission for predicting length of stay (LoS). Mean with 95% CIs is calculated using bootstrapping. AUPRC, area under the precision recall curve; AUROC, area under the receiver operating curve.

• 

  Figure 4. Feature importance from the extreme gradient boosting models predicting in-hospital mortality and future hospital length of stay (LoS) using SHapley Additive exPlanations values for elective and emergency admissions. The top 20 most predictive features are shown in the order of predictiveness. APTT, activated partial thromboplastin time; BMI, body mass index; CRP, C-reactive protein.

• 

  Figure 5. Comparison between clinician predictions and model predictions on future hospital length of stay (LoS) in the test dataset where clinicians made predictions. (a) Comparison of accuracy, precision, recall and F1 score between clinician and the best-performing extreme gradient boosting (XGB) model for elective admissions. (b) Comparison of accuracy, precision, recall and F1 score between clinician and the best-performing XGB model for emergency admissions. (c) Confusion matrix for clinician predicted and actual LoS quartiles for elective admissions. (d) Confusion matrix for clinician predicted and actual LoS quartiles for emergency admissions. (e) Confusion matrix for model predicted and actual LoS quartiles for elective admissions. (f) Confusion matrix for model predicted and actual LoS quartiles for emergency admissions. Q1, below 25% quantile (0–1 days for elective, 0–2 for emergency); Q2, 25% to 50% quantile (2–3 days for elective, 3–6 days for emergency); Q3, 50% to 75% quantile (4–10 days for elective, 7–14 days for emergency); and Q4, above 75% quantile (>10 days for elective, >14 days for emergency). Means with 95% CIs are calculated using bootstrap.