Objective: To detect real-life tremor and dyskinesia fluctuations, using naturalistic wearable data for prediction, and ecological momentary assessments (EMA) data for model training.

Background: Parkinson’s disease (PD) patients suffer from motor fluctuations, impairing their quality of life. Deep brain stimulation (DBS) improves motor fluctuations and dopaminergic side effects, such as “wearing-OFF”, tremor, and levodopa-induced dyskinesia (LID). To reliable capture motor symptoms and DBS effect in real-life, symptom prediction needs to apply reliable naturalistic ground truths.

Method: First, a cohort of 24 PD patients completed a PD-specific EMA questionnaire and where assessed using the UPDRS III scale, in four different therapeutic state, i.e. with/without dopaminergic medication, and with/without subthalamic DBS. Second, we included two subjects as use-cases to showcase EMA-trained symptom detection models based on sensor-derived features. Subject A, a tremor-dominant Parkinsonian patient, and subject B, a PD patient with LID, collected EMA- and sensor-data during 14 days, both pre-DBS and post-DBS. Priory published accelerometer-features associated with tremor and LID were extracted. For subject A, we performed an unsupervised clustering to differentiate between tremor-severities and for subject B we performed a supervised cross-validation classification to predict LID severities.

Results: We observed strong correlations between individual motor fluctuations captured with EMA and UPDRS-III assessments for the three motor domains tested: bradykinesia: pearson-r = 0.72, tremor: pearson-r = 0.72, and gait: pearson-r = 0.56 (all p-values below 0.0001). In the tremor use case, a combined RandomTree and Clustering method generated three clusters, of which the sorted mean EMA-tremor values correlated significantly (spearman rho 0.40, permuted p-value: 0.01). In the cross-validation with preoperative data of the LID use case, sensor-derived features predicted LID-answers (spearman rho 0.46, p<0.001).

Conclusion: We validated EMA self-report by correlating EMA-motor domains with UPDRS scores. Moreover, we showed two naturalistic use cases of sensor-based symptom detection, trained or validated with EMA.

« Back to 2026 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/real-life-accelerometer-based-detection-of-dbs-effects-on-tremor-and-dyskinesia-in-parkinsons-disease-trained-with-ecological-momentary-assessments/