PathoFM: Toward a Foundation Model for Pathological Gait

Pathological gait exhibits diverse compensatory strategies that vary across individuals, disease stages, and time. Robust downstream clinical performance can benefit from foundation models that learn generic, transferable motion representations rather than task-specific features. However, an interesting question is what inductive biases prove to be good training objectives for a general recipe to train such FMs. We address this with PathoFM, an encoder-only pretraining recipe trained on heterogeneous gait cycles from 230 patients, augmented with synthetic generative variants of real trials to broaden coverage of atypical patterns. The recipe blends three complementary objectives: (i) Local Completion (recovering continuous segments of the input), (ii) Temporal Continuity (predict future segments to enforce dynamic consistency), and (iii) In-Context Dynamics, an unsupervised in-context learning objective that encourages relational reasoning from a small support set of exemplars. We evaluate under strict patient (subject) holdout and compare PathoFM against grouping-based pretexts (subject-ID discrimination, InfoNCE contrastive learning, online prototypes) and diffusion variants. Across clinical classification and regression endpoints, PathoFM achieves the best overall balance of performance. These results indicate that dynamics-centric pretraining yields more generalizable clinical timeseries representations than objectives based on grouping or instance discrimination.