Methods for tracking lab animal movements in unconstrained environments have become increasingly common and powerful tools for neuroscience. The prevailing hypothesis is that animal behavior in these environments comprises sequences of discrete stereotyped body movements ("motifs" or "actions"). However, the same action can occur at different speeds or heading directions, and the same action may manifest slightly differently across subjects due to, for example, variation in body size. These and other forms of nuisance variability complicate attempts to quantify animal behavior in terms of discrete action sequences and draw meaningful comparisons across individual subjects. To address this, we present a framework for motion analysis that uses conditional variational autoencoders in conjunction with adversarial learning paradigms to disentangle behavioral factors. We demonstrate the utility of this approach in downstream tasks such as clustering, decodability, and motion synthesis. Further, we apply our technique to improve disease detection in a Parkinsonian mouse model.

Early and accurate triage of traumatic brain injury is critical for guiding treatment decisions that optimize patient outcomes. A major early clinical decision point occurs in the emergency department, where providers must decide whether to admit or discharge patients with head injuries, yet these decisions are often inconsistent and rarely supported by case-based frameworks. Here, we introduce RAPID-TBI (Retrieval Augmented Prediction for Informed Disposition in Traumatic Brain Injury), a multimodal system that predicts emergency department disposition using example-based retrieval to emulate clinical case-based reasoning. RAPID-TBI achieves state-of-the-art classification performance while enhancing interpretability by retrieving similar patients to inform predictions. Using a large multimodal TBI dataset from a major U.S. hospital system, RAPID-TBI integrates head CT scans, radiology reports, exam findings, laboratory values, vitals, and demographics through an attention-based encoder that generates patient embeddings for disposition classification. We further assessed RAPID-TBI across institutional and temporal generalizability, showing consistent performance and resilience to shifts in data distribution. Finally, we explored small language models as prompt-based classifiers for retrieval-guided prediction without fine-tuning. Together, these components enable RAPID-TBI to deliver consistent, individualized, and clinically grounded predictions, a promising step toward trustworthy, personalized decision support in TBI care.

Parkinson's disease (PD) is a devastating movement disorder accelerating in global prevalence, but a lack of precision symptom measurement has made the development of effective therapies challenging. The Unified Parkinson's Disease Rating Scale (UPDRS) is the gold standard for assessing motor symptom severity, yet its manual scoring criteria are vague and subjective, resulting in coarse and noisy clinical assessments. Machine learning approaches have the potential to modernize PD symptom assessments by making them more quantitative, objective, and scalable. How-ever, the lack of benchmark video datasets for PD motor exams hinders model development. Here, we introduce the TULIP dataset to bridge this gap. TULIP emphasizes pre-cision and comprehensiveness, comprising multi-view video recordings (6 cameras) of 25 UPDRS motor exam activities, together with ratings by 3 clinical experts, in a cohort of Parkinson's patients and healthy controls. The multi-view recordings enable 3D reconstructions of body movement that better capture disease signatures than more conventional 2D methods. Using the dataset, we establish a base-line model for predicting UPDRS scores from 3D poses, illustrating how existing diagnostics could be automated. Looking ahead, TULIP could aid the development of new precision diagnostics that transcend UPDRS scores, providing a deeper understanding of PD and its potential treatments.