Objective: To develop a classification algorithm based on machine learning to identify candidates for device-aided therapy among patients with Parkinson’s disease (PD).

Background: DATs (deep brain stimulation and continuous apomorphine and levodopa infusions) are indicated in patients with intermediate/advanced PD when conventional treatment fails to control disabling motor fluctuations, dyskinesias, and/or tremor. Tools have been developed to facilitate the identification of patients with fluctuations (e.g. 5-2-1) and candidates for DAT (e.g. MANAGE PD) based on expert consensus. Artificial intelligence using machine learning algorithms developed with real cases could offer a promising tool to facilitate identification for DAT.

Method: This is a national, multicenter, cross-sectional study in PD patients recruited during 12 months in 9 Movement Disorders Units according to cluster sampling (25% candidates for DAT regardless of the stage of the disease; 50% initial or intermediate stage and 50% advanced). Gold standard for DAT candidate identification was the expert neurologist’s decision. Predictive variables gathered were, among others: time in off/on and with/without disabling dyskinesia, doses of oral levodopa/day, Charlson index,  quality of life (PDQ-8), cognitive impairment (GDS) and CDEPA Questionnaire. Catboost with stratified cross-validation 10 folds and balance techniques were applied to confirm discriminative capacity and improve minority class detection. The relative influence of variables was estimated with Sequencial Forward Selection.

Results: We studied 1086 PD patients (69.6±10.5 yrs, 57% male) with median H&Y scale of 2 and mean duration of 8.2±5.7 yrs.  Patients identified as advanced PD were 43%, 194 (42%) of them candidates for DAT. The maximum value of balanced accuracy was obtained with 23 variables (90% of the value); the most influential were daily OFF time (>80% of the value), doses of oral levodopa/day and disease duration.  Accuracy of catboost identification was 89% (Sensitivity=0.91±0.03; Specificity=0.88±0.05). AUC and precision-recall curves was 95%.

Conclusion: This algorithm based on real cases from the clinical practice showed good discriminative capacity to detect PD candidates for DAT and may be a useful tool for general neurologists. The results were consistent with key criteria from previous expert consensus.

« Back to 2023 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/predictive-model-for-identification-of-device-aided-therapy-dat-candidates-in-parkinsons-disease-delist-pd-study/