Personalized medication recommendations for Parkinson's disease patients using gated recurrent units and SHAP interpretability

Abstract

Managing Parkinson's disease (PD) through medication can be challenging due to varying symptoms and disease duration. This study aims to demonstrate the potential of sequence-by-sequence algorithms in recommending personalized medication combinations for patients with PD based on their previous visits. Our proposed method employs a gated recurrent unit model to predict accurate combinations of critical medication types for PD based on each patient's motor symptoms and prescribed medication from previous visits. We built a multi-label model with gated recurrent units on two data architectures: (1) personalized input using each patient's previous visits as a sample and (2) non-personalized input treating each visit as an independent sample. The 10-fold cross-validation results showed that the personalized architecture model outperforms the non-personalized model in accuracy (0.92), precision (0.94), recall (0.94), F1-score (0.94), Hamming loss (0.03), and macro average area under the receiver operating characteristic (0.94). To interpret the model's predictions, we employed SHapley Additive exPlanations (SHAP) values, which provide insights into the importance of variables both globally (across the entire model) and at the individual patient level. The results contribute to the sequential-based decision support system potentially enhancing the remote management of PD pharmacologic issues.

Keywords: Decision support systems; Explainable artificial intelligence; Personalized decisions; Pharmacological decisions; Sequential learning.

Fig. 1

Different combinations of LD (L), DA (D), and Other (O) medication, including MAOBi, COMTi, and Amantadine. There are 8 possible combinations with three types of levodopa (L), DA (D), and Other medications (O), from No medication (NoMed), to LD + DA + Other (LDO).

Fig. 2

The personalized model architecture for the medication combination prediction. The model processes twelve features over five time-steps, utilizing two GRU layers and a fully connected dense layer, with a Sigmoid output for multi-label medication prediction.

Fig. 3

The ROC of one vs. rest classification of personalized and non-personalized multi-label predictions. (a) Personalized input with macro-average and micro-average AUC of 0.94 and 0.96, respectively, showing higher AUC for individual medication classes. (b) Non-personalized input with macro-average and micro-average AUC of 0.85 and 0.84, demonstrating lower performance compared to the personalized model.

Fig. 4

The importance of input features in predicting medication.

Fig. 5

SHAP local interpretation for a sample patient’s medication usage: (a) LD, (b) DA, and (c) other medications. The SHAP values, averaged over the last five visits, highlight the positive and negative impacts of features such as previous medication usage, age, and symptoms like tremor and bradykinesia on medication probabilities.