Time-Series Patient-Reported Data and LSTM Predicting Hospital Utilisation During Chemotherapy

Adverse effects of chemotherapy often require acute hospital admissions, which can negatively impact patients’ well-being and increase healthcare burden. Identifying the risk of hospital utilisation could support prevention of patients’ deterioration and alert medical teams about potential admissions. This study uses patients’ clinical and demographic data, patient-reported outcome measures, and time-series symptom severity reports to predict the risk of hospital admissions and triage events within 14 days from completed symptom severity report. Hospital utilisation at any time during chemotherapy was also predicted. The performance of long-short term memory (LSTM) and extreme gradient boosting (XGBoost) models was compared. Nested cross-validation enabled robust hyperparameter tuning and model evaluation using unseen data. Patient representatives and a clinical oncologist were consulted during the study design to support its clinical relevance. LSTM outperformed XGBoost at short-term predictions of hospital admission (balanced accuracy=0.780, AUC=0.845) and triage (balanced accuracy=0.706, AUC=0.779). However, XGBoost performed better at long-term predictions. The results suggest that LSTM processed complex data with sudden fluctuations better than XGBoost. However, classical ML might be sufficient for longer-term outcome predictions. If further explored, these models could prompt hospital contact prior to patient’s deterioration and prevent admission or alert medical team of potential hospitalisation.