medX

Inspiration

Healthcare AI can learn from enormous amounts of clinical data, but that data is fragmented across hospitals and cannot simply be centralized because of privacy, security, and governance concerns.

Federated Learning solves part of this problem by allowing hospitals to collaboratively train models without directly sharing raw patient records.

But we found a second, equally important problem:

Even if hospitals can safely learn together, should that shared intelligence be trusted for every patient?

A federated model may encounter patients poorly represented across participating hospitals. Hospital populations can shift. Data can become unreliable. A model can become uncertain while still producing a confident-looking prediction. Improvements in overall performance can also hide degradation for smaller or underrepresented patient groups.

That led to MedX.

MedX asks not only what the AI predicts, but whether its federated knowledge should be trusted for this patient, at this hospital, at this moment.

---

What it does

MedX is a patient-aware federated clinical intelligence system designed around patient-time assurance.

It combines federated learning with four core capabilities:

Guardian

Evaluates whether federated knowledge is appropriate for the current patient using representation, distribution and uncertainty signals.

Guardian can return:

• ACCEPT — sufficient support exists
• CAUTION — additional clinical attention is required
• ABSTAIN — AI recommendation is withheld for human-only assessment

Sometimes the safest AI prediction is no prediction at all.

Sentinel

Monitors participating hospitals and helps distinguish between:

• Legitimate clinical or population drift
• Data-quality problems
• Unreliable model updates
• Potentially harmful behaviour

MedX can then ADAPT, REVIEW, or REJECT an update instead of treating every unusual hospital as untrustworthy.

Safety Passport

Instead of showing clinicians only a prediction such as "CKD Risk: 82%", MedX attaches evidence explaining the prediction's reliability.

The Safety Passport can include:

• Patient representation
• Predictive uncertainty
• Cross-hospital support
• Distribution-shift status
• Local validation status
• Model provenance and version

EquityGuard

Monitors whether improvements to the federated model are hiding performance degradation for smaller hospitals, rare clinical cohorts, or underrepresented patient populations.

Together, these capabilities move federated healthcare from training-time privacy to patient-time trust.

---

How we built it

MedX builds on our existing FedMed federated healthcare foundation.

Each participating hospital keeps its patient data locally and performs local model training. Permitted model information can then contribute to the federation without creating a centralized repository of raw patient records.

Our foundation uses:

• Python
• PyTorch
• Flower Federated Learning
• FastAPI
• Non-IID multi-hospital datasets
• Local clinical inference
• Human clinical oversight

The proposed MedX assurance architecture extends this foundation:

Hospital A ─┐
Hospital B ─┼──> FedMed Core
Hospital C ─┘        |
                     v
                  Sentinel
                     |
                     v
            Federated Clinical Model
                     |
Patient ─────────> Guardian
                     |
          +----------+----------+
          |          |          |
       ACCEPT     CAUTION    ABSTAIN
          |          |          |
          +----------+----------+
                     |
                     v
              Safety Passport
                     |
                     v
                  Clinician

             <--- EquityGuard --->

Built With

• ai
• fastapi
• fl
• flower
• ml
• next.js
• python
• pytorch
• react
• sqlite
• typescript