Recurrent neural networks and attention scores for personalized prediction and interpretation of patient-reported outcomes.

Abstract

We proposed an Interpretable Personalized Artificial Intelligence (AI) model for PRO measures via Recurrent Neural Networks (RNN) and attention scores, with data from an open label randomized clinical trial of pain in 402 participants with cryptogenic sensory polyneuropathy at 40 neurology care clinics. All patients were assigned to four treatment groups: nortriptyline, duloxetine, pregabalin, and mexiletine. Each patient had 4 PRO measures (quality of life SF-12; PROMIS: pain interference, fatigue, sleep disturbance) at 4 time points (baseline, week 4, week 8, and week 12). We included 201 patients without missing values. Participants were 30 years or older and 106 (52.7%) were men, majority were White (164, 81.6%). We fitted an RNN model with attention scores to the data to predict the PROMIS pain interference score. We evaluated the model performance with Mean Absolute Error (MAE) and R-square statistics. We also used attention scores to explain the global variable importance at model level, and at individual level for each patient. The best predictor of pain score was the SF-12 item (physical and emotional health interfere with social activities) and fatigue item (push yourself to get things done), the biggest drug-level contributor was mexiletine, the biggest time-level contributor was week 12. Overall, the model fit well (MAE = 3.7, R2 = 63%). Attention-RNN is a feasible and reliable model for predicting PRO outcomes utilizing longitudinal data, such as pain, and can provide personalized individual level interpretation.