NeuroLynk AI

• 

  The screening interface allows adjustment of individual vocal biomarkers with real-time SHAP attributions.

• 

  Generates a clinician-readable narrative of the findings using Large Language Models.

• 

  Global feature importance and comparative insights derived from the training distribution.

Inspiration

Parkinson's disease affects over 10 million people worldwide, yet early diagnosis remains elusive. Studies show that vocal biomarkers — subtle changes in pitch, tremor, and spectral energy — can precede motor symptoms by years. But these signals are invisible to the human ear. I was inspired by the question: What if an AI agent could listen to speech and flag early Parkinson's risk before a patient even walks into a clinic? More importantly — what if it could explain why it made that decision, in language a clinician could trust?

What it does

NeuroLynk-AI is a healthcare AI agent for explainable Parkinson's speech screening. It analyzes 753 vocal biomarkers extracted from speech recordings (MFCCs, delta coefficients, formants, log energy, wavelet decomposition levels) and returns a full clinical screening result.

• Prediction: Healthy or Parkinson's Detected with confidence percentage
• Explainability: Top contributing biomarkers identified via SHAP values
• Clinical Summary: Plain-language interpretation generated by Gemini 2.5 Flash
• FHIR R4 Report: Standards-compliant DiagnosticReport for EHR integration
• A2A Ready: Fully registered on Prompt Opinion as a JSON-RPC agent, callable by other agents

How we built it

The SHAP value $\phi_i$ for each biomarker feature $i$ is computed as: $$\phi_i = \sum_{S \subseteq F \setminus {i}} \frac{|S|!(|F|-|S|-1)!}{|F|!} \left[ f(S \cup {i}) - f(S) \right]$$ This guarantees additive feature attribution — every prediction is a transparent sum of individual biomarker contributions.

Challenges we ran into

• A2A Schema Compatibility: The Prompt Opinion validator enforced a strict AgentCardV1 schema. Aligning protocolBinding: "JSONRPC", text/plain modes, and empty securitySchemes: {} required significant iteration.
• SHAP Version Conflicts: Different SHAP versions return values as list, 2D array, or 3D array. We built a robust normalisation function to handle all output shapes.
• LLM Environment Loading: On Cloud Run, python-dotenv was not loading .env at startup, causing the Gemini provider to silently fall back to mock mode. Fixed by calling load_dotenv() explicitly at the top of main.py.
• Wavelet Feature Interpretation: Wavelet coefficients at different decomposition levels represent different frequency bands and are not directly comparable in magnitude, requiring careful normalization before model input.

Accomplishments that we're proud of

• Achieved ~94% accuracy and 0.97 ROC-AUC on the Oxford Parkinson's dataset
• Built a fully standards-compliant FHIR R4 DiagnosticReport generator
• Successfully registered as a live A2A agent on the Prompt Opinion platform with JSON-RPC 2.0 compliance
• Designed a production-ready system with MLflow experiment tracking and KS-test drift detection for model monitoring

What we learned

• The A2A protocol requires a precise agent card schema — small field name or type mismatches cause silent 422 validation failures
• Explainability is not optional in healthcare AI; SHAP makes the difference between a model a clinician ignores and one they trust
• FHIR R4 is a powerful but strict standard; even minor structural deviations break EHR compatibility
• Cloud Run environment variable propagation requires explicit handling — never assume .env is loaded automatically

What's next for NeuroLynk AI

• Real audio pipeline: Accept raw .wav files and extract features on-the-fly instead of requiring pre-computed vectors
• Longitudinal tracking: Monitor a patient's vocal biomarker trends across multiple screenings over time
• Multi-disease extension: Adapt the pipeline to screen for ALS and essential tremor using the same speech biomarker framework
• EHR Integration: Connect directly to FHIR servers for automated report submission and patient record updates
• Clinical trial partnership: Validate NeuroLynk-AI against prospective clinical data for regulatory pathway exploration

Built With

• a2a-protocol
• fastapi
• fhir
• gemini
• json-rpc-2.0
• mlflow
• prompt-opinion
• pydantic
• python
• scikit-learn
• shap
• xgboost