Development and validation of a machine learning model to detect psychiatric symptoms in Huntington’s disease using speech analysis

Huntington's disease (HD) causes progressive disability through motor, psychiatric, and cognitive symptoms. Machine learning speech analysis can detect motor and cognitive symptoms of HD, but not yet psychiatric symptoms. This study investigated whether speech analyses can detect the presence of psychiatric symptoms in HD. Audio recordings of six narrative tasks (cookie-theft picture description, red-riding hood storytelling, most recent 24 hours recalling, happy, sad, or angry storytelling) were prospectively collected from subsequent genetically confirmed HD participants from the BIOHD and REPAIR CAPIT-HD-Beta cohorts at the Hospital Henri-Mondor, Créteil. Speech therapists blindly annotated speech samples to allow extraction of three types of features: linguistic, LASER, and acoustic features. Psychiatric symptoms in participants were detected using the Problem Behaviors Assessment Short version (PBA-s). Machine learning classifier models were trained on 80% of the 89 participants before being tested on the remaining 20% of individuals. F1-scores were calculated and compared to chance. Linguistic features detected obsessive/compulsive behavior (OCB) with all but joy task, and best with the cookie task (F1-score: 0.67, confidence interval [0.47-0.86] (p ≦ 0.001)). They also best detected depression with the red-riding hood (F1 score 0.66, [0.45-0.87], p ≦ 0.001), apathy with the joy task (0.60, [0.39-0.81], p ≦ 0.001), but not irritability. LASER features best detected OCB (0.65, [0.45-0.84], p ≦ 0.001), depression (0.60, [0.40, 0.80], p ≦ 0.01) and apathy (0.61, [0.37, 0.86], p ≦ 0.001) from the red-riding hood task, but not irritability. Acoustic features best detected depression (0.63, [0.46, 0.80], p ≦ 0.001) and OCB (0.60, [0.43, 0.77], p ≦ 0.001) but not apathy nor irritability. This study showed that speech analyses can detect obsessive/compulsive behaviors, depression, and apathy in HD participants but not irritability. Linguistic and LASER features provided the most consistent detections, but acoustic features also detected depression and OCB, highlighting their complementary role for psychiatric characterization in HD.