TriAD is a clinical-grade frontend interface designed to screen for Alzheimer's Disease (AD) by triangulating digital biomarkers with proprietary genetic and structural datasets.

System Status: Data Ingestion & Validation Mode

This platform serves as the deployment vehicle for our custom-trained risk models. It is architected to capture real-time user telemetry (voice acoustics, oculomotor latency) and correlate it with our ground-truth clinical cohorts.

Unlike generic screening tools, TriAD is built to interface directly with our validated clinical datasets:

1. Genetic Knowledge Base (advp.hg38.tsv):

   • Role: Explainable AI (XAI) Dictionary.
   • Function: Maps 130+ detected SNPs to clinically validated risk ratios (Odds Ratios), allowing the system to explain why a genetic flag was raised (e.g., "Detected rs429358 on APOE gene").

2. Patient Genetic Data (preprocessed_alz_data.npz):

   • Role: Training ground for Agent 2.
   • Scale: 5,076 patients with 130 preprocessed genetic features.
   • Model: Random Forest classifier calibrated to detect "Silent Risk" before symptoms appear.

3. Imaging Data (train.parquet):

   • Role: Validation set for Agent 3.
   • Content: MRI scans labeled by severity (Non-Demented vs. Mild Demented).
   • Model: ResNet/CNN architecture for structural segmentation and atrophy detection.

The system utilizes a multi-agent "Sensor & Inference" architecture:

Status: Active Telemetry Implementation Because cognitive decline is dynamic, Agent 1 acts as a Real-Time Sensor to collect fresh digital biomarkers that standard datasets lack.

• Acoustic Feature Extraction: Uses the Web Speech API to capture natural language samples, analyzing hesitation rate and vocabulary richness (Temporal Lobe indicators).
• Inhibition Latency: Uses high-precision performance.now() timers in the interactive Stroop instrument to measure processing speed (Frontal Lobe indicators).

Status: Model Integration Phase

• Input: User-reported status or VCF file ingestion.
• Logic: Runs inference against the preprocessed_alz_data.npz model.
• Explainability: Cross-references positive hits against advp.hg38.tsv to provide medical context rather than black-box scores.

Status: Imaging Pipeline

• Input: DICOM/Image upload pipeline.
• Logic: Pre-processing interface for ResNet/CNN inference models trained on train.parquet.

• Phase 1 (Current): Deploy "Sensor" frontend (Agent 1) to capture standardized inputs (Voice/Reaction Time).
• Phase 2 (Training): Train the Genetic Model (Random Forest) and Imaging Model (FastAI/PyTorch) using the provided datasets.
• Phase 3 (Integration): Connect the frontend risk.ts scoring logic to a Python backend serving the trained models.

• Frontend: React + TypeScript, Vite, Tailwind CSS
• Sensors: Web Speech API, High-Resolution Time API
• Visualization: Recharts, Framer Motion
• Planned Backend: Python (FastAPI), Scikit-learn (Genetic), PyTorch (Imaging)

This software is a research instrument designed for use with specific clinical datasets. It is not a standalone diagnostic device. Always consult qualified healthcare professionals for medical advice, diagnosis, and treatment decisions.