Skip to content

Sonata: A Hybrid World Model for Inertial Kinematics under Clinical Data Scarcity

Home / Papers / Sonata: A Hybrid World Model for Inertial Kinematics under Clinical Data Scarcity

Sonata: A Hybrid World Model for Inertial Kinematics under Clinical Data Scarcity

Authors: Blaise Delaney, Salil Patel, Yuji Xing, Dominic Dootson, Karin Sevegnani Date: 2026-04-20 Paper ID: openalex:2604.18058

Summary

Sonata is a compact, hybrid latent world model designed for six-axis trunk IMU representation learning, specifically tailored for scenarios with limited clinical data. Unlike standard masked-reconstruction objectives, Sonata utilizes a predictive future-state objective to learn structured representations from a harmonized corpus of inertial sensor data. Empirical results across 14 evaluation tasks demonstrate that this approach consistently outperforms autoregressive forecasting baselines in clinical discrimination and fall-risk prediction tasks. The model’s small parameter footprint facilitates efficient on-device inference for neurological assessment.

Key Contributions

  • Introduces Sonata, a 3.77M-parameter hybrid latent world model for six-axis trunk IMU representation learning.
  • Demonstrates superior clinical discrimination and fall-risk prediction performance compared to masked-reconstruction (MAE) baselines under data-scarce clinical conditions.
  • Achieves effective cross-cohort transfer learning across a 14-arm evaluation suite while remaining compatible with on-device wearable inference.

Open Questions & Future Work

Key Concepts

  • latent-world-model-kinematics: A representation learning approach that uses predictive future-state objectives rather than raw signal reconstruction to capture structured kinematic dynamics in low-data regimes.

Archivist Review

I approved the latent world model for kinematics concept as it captures the specific shift towards predictive future-state objectives over masked-reconstruction in data-scarce time-series domains. The open questions regarding longitudinal clinical monitoring and multi-site kinematic modeling were selected because they identify fundamental bottlenecks in deploying wearable foundation models in clinical settings beyond this specific study. Sonata itself was rejected as a concept to keep the focus on the broader mechanism.

Approved Concepts

  • Latent World Model for Kinematics: Represents a shift from standard masked-reconstruction tasks to predictive future-state tasks in small-data inertial time series, offering a more robust representation for clinical downstream tasks.

Approved Open Questions

  • Longitudinal Clinical Monitoring via Wearables: Infrequent clinical rating scales are insensitive to subtle fluctuations; developing a longitudinal methodology for wearables is essential for clinical trial design and neurological patient management.
  • Multi-site Kinematic World Modeling: Relaxing placement constraints would allow for larger training corpora and more comprehensive whole-body representations necessary for complex neurological movement disorders.

Rejected Candidates

  • [concept] Sonata (Hybrid World Model) (sonata-model) - subcomponent_of_broader_mechanism: The model itself is a specific architectural implementation; the underlying approach is better represented by the concept of the latent world model for kinematics.

Metadata & Links

url
https://arxiv.org/abs/2604.18058
paper_id
2604.18058
paper_source
openalex
domain
time-series
tags
architectures
datasets
concept_slugs
latent-world-model-kinematics
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-04-23T05:48:57Z