Machine Learning Models Predicting Hospital Admissions During Chemotherapy Utilising Longitudinal Symptom Severity Reports and Patient-Reported Outcome Measures

Chemotherapy toxicity can lead to acute hospital admissions, negatively impacting the healthcare system and patients' well-being. Machine learning (ML) models identifying patients at risk of emergency admissions are often developed on data lacking patients' perspective. This study used longitudinally collected symptom severity reports and 4 ML models to predict hospital admissions risk during chemotherapy, and short-term admissions risk (within 14 days of a report). It also compared performance of models developed with, and without the use of patient-reported outcome measures (PROMs). Random forest and extreme gradient boosting models predicted admissions with excellent balanced accuracy, recall, and specificity of over 0.9. However, short-term admissions risk predictions were poor. PROMs improved overall model performance. The results advocate for longitudinal collection and use of symptom severity reports and PROMs. This can support understanding of chemotherapy toxicity patterns leading to emergency admissions, and inform clinicians and patients of potential future complications.