Using Machine Learning for Predicting Non-Adherence in Remote Oncology Symptom Monitoring

Electronic patient-reported symptom monitoring helps manage symptoms in cancer care, but patients do not always report every day. We aim to predict short-term non-adherence so teams can reach out before monitoring breaks down. We treat this as a next-window forecast using overlapping 3-, 5-, and 7-day windows of recent symptoms. We create one record per patient-day, rebuild missing days, apply simple inclusion rules (≥35 observed days and ≤30% missing), and fill partial gaps within each window by averaging the days that are present (set to 0 only if all days are missing). We label each window by whether the following window has any missed day. To handle imbalance, we use class-weighted loss. We test four models: LSTM, CNN-LSTM, CNN-(Bi)LSTM with single-head attention, and CNN-(Bi)LSTM with multi-head attention. We standardize features, use an 80/20 stratified split. We report ROC-AUC, accuracy, precision, recall, F1, and confusion matrices. With 7-day windows, LSTM and CNN-LSTM with attention give the best ranking (ROC-AUC ≈ 0.755) and balanced accuracy (≈0.71-0.73). CNN-LSTM reaches high recall (class-1 recall 0.90) but lowers accuracy (0.56). With 5-day windows, CNN-LSTM gives the highest positive-class F1 at its tuned threshold (validation recall ≈0.93; held-out recall 0.46), making it strong for catching imminent lapses. With 3-day windows, tabular baselines (XGBoost) raise accuracy (≈0.75) but have lower F1 and AUC than the best 5-/7-day sequence models. In sum, short-horizon non-adherence is predictable from recent ePRO patterns.