Clinical and digital‑health models must integrate data from many heterogeneous and rapidly changing sources (EHR systems, devices, apps), making conventional models fragile when inputs or formats change.​ Maintaining and re‑validating separate models for each data source or version is costly and slow, limiting deployment of AI‑based decision support at scale.​

The technology

• Autoencoder modules trained to receive “content data elements” about a subject from multiple data sources and convert them into a feature vector
• ML classification models trained to take the vector and output a predicted condition of the subject (e.g., disease state, risk, or other target)
• Dedicated training modules orchestrate a multi‑stage process in which autoencoders and classifiers are trained on annotated data while discarding source‑specific artifacts​

Potential applications

• Clinical decision‑support tools
• Digital‑health platforms that combine data from multiple apps, wearables, or remote‑monitoring devices while insulating the core models from frequent changes in vendors and data formats
• Any multi‑source analytics system (such as population‑health, RWE, or behavioral analytics) that requires stable predictions despite evolving data feeds

Reference: 

US2024/0104350 a1 – “machine learning‑based invariant data representation”: https://patents.google.com/patent/US20240104350A1