Sensor-based data-driven differentiation between Parkinson’s tremor and essential tremor

• Tremor differentiation model validated across hospitals and equipment, demonstrating real-world clinical generalizability. • Multi-modal EMG/accelerometer framework: 98 % internal accuracy, 79 % external test ROC-AUC. • Practical expert system ready for clinical deployment with minimal preprocessing and robust performance across recording conditions. Despite distinct pathophysiologies, tremor due to Parkinson’s disease and essential tremor are commonly misdiagnosed due to overlapping symptoms, hindering clinical intervention. Digital phenotyping with machine learning applied to peripheral sensor-based signals has shown potential. However, it lacks validation using an independent dataset or commercial medical device recordings broadly available for routine clinical use. Here, we present a scalable and generalizable diagnostic approach using features engineered from frequency, power, and non-linear phase domains combined with tree-based classification algorithms. Using data from a single hospital, our XGBoost model achieved a diagnostic specificity of 0.92, recall of 1.0, and ROC-AUC (Receiver Operating Characteristic − Area Under the Curve) of 1.0 on a held-out test set. Applying our pre-trained model to an external dataset, recorded in a different clinical setting using other devices, demonstrated strong generalizability with specificity of 0.70, recall of 0.80, and ROC-AUC of 0.79. Our findings not only outperform previous studies in predictive accuracy but also deliver clinically validated metrics that can be implemented in practice. These results highlight the feasibility of deploying affordable, widely available sensor-based diagnostics to enhance clinical accuracy and inform adaptive therapeutic interventions.