Developing machine learning-enhanced WHODAS 2.0 short forms for persons with dementia

Disability assessment in dementia is important for care planning, but the full World Health Organization Disability Assessment Schedule 2.0 (WHODAS 2.0) is time-intensive and may limit clinical use. This study developed machine learning (ML)-based short forms of the WHODAS 2.0 and examined their reliability, concurrent validity, and responsiveness.

Using data from 51,245 persons with dementia (training set: n = 31,952; validation set: n = 19,293), we developed two ML-based short forms, ML-WHODAS-16 and ML-WHODAS-10, with Extreme Gradient Boosting and bootstrap-based item selection under a lock-down training/validation workflow. Their performance was compared with the full WHODAS-32 and the conventional 12-item short form adapted to 11 items after excluding work-related items (Standard-12). Anchor-based longitudinal validity was also examined using deterioration in official Disability Severity Grade.

Both ML short forms showed high internal consistency (α = 0.96 for ML-WHODAS-16 and 0.93 for ML-WHODAS-10) and excellent concurrent validity with the full WHODAS-32 (r = 0.98 for both). Compared with the Standard-12, they showed lower error, negligible Bland-Altman bias, and met predefined equivalence criteria, including ±0.5 points. Anchor-based Responsiveness was broadly comparable to the Disability Severity Grade (anchor) (r = 0.66-0.67; standardized response mean = 0.37-0.40). Anchor-based minimal clinically important differences were 9.26 for ML-WHODAS-16 and 9.95 for ML-WHODAS-10.

The ML-WHODAS-16 and ML-WHODAS-10 substantially reduced assessment burden while maintaining scores that closely reflected those of the full WHODAS-32, particularly for group-level assessment and longitudinal monitoring. These findings support their use as practical, low-burden alternatives in dementia disability assessment. However, external validation, validation against harder clinical outcomes, formal non-inferiority testing, and clinically anchored longitudinal thresholds remain needed before individual-level interchangeability can be inferred.