Yu, X.; Martens, K. E.; Arami, A.

Freezing of gait (FOG), a disabling symptom of Parkinson's disease, varies in presentations and motion context. Its heterogeneity motivates subtype categorization such as manifestation-specific subtypes (akinesia, trembling or shuffling) or motion-specific subtypes (gait-initiation, walking or turning). Despite numerous deep learning FOG detection models, few consider FOG heterogeneity. Thus, it remains unclear whether subtype-specific models could enhance detection generalizability. This paper clusters FOG data into manifestation- or motion-specific subtypes and encodes their detection strategies in feature masks. It is found that while motion-specific subtypes share a common detection strategy, manifestation-specific subtypes require distinct strategies. A feature-mask-based CNN is proposed to train general and subtype-specific models using waist-mounted 3D accelerometer data. Manifestation models exhibit enhanced generalizability across subtypes over a general model, boosting the overall average FOG sensitivity by 24.95% +/- 9.80 and non-FOG specificity by 18.29% +/- 8.71. Differently, motion models show no such improvement. This suggests that detection strategy is driven by manifestation composition of data. The general model favors the dominant manifestation-specific subtype(s), a bias corrected by tailored manifestation-specific strategy. No comparable benefit arises from motion models due to their similar manifestation compositions.