AI-Driven Diagnosis Accelerates Undiagnosed Rare Syndromes
A collaboration between Toronto's SickKids, São Paulo's Genomic Center, and Tokyo Medical AI Lab demonstrates how multimodal machine learning fast-tracks diagnoses for children with unresolved rare phenotypes.
Model Architecture
The ORIGIN platform merges facial phenotyping, clinical notes embeddings, and trio exome sequencing data. A transformer-based fusion layer prioritizes candidate diagnoses.
Performance
Top-5 diagnosis accuracy reached 84% across 1,100 cases, cutting time-to-diagnosis from 24 months to 7 months.
Bias Mitigation
Model retraining with diverse phenotypic datasets reduced accuracy gaps across ancestry groups from 18% to 5%.
Clinical Integration
Clinicians upload patient data through a secure portal. AI-generated differential diagnosis lists align with HPO terms and suggest confirmatory tests.
- • Integration with telehealth consults for remote triage.
- • Automated consent workflows for data sharing across borders.
- • API endpoints enabling linkage to national newborn screening registries.
Pilot Hospitals
- SickKids, Toronto, Canada
- Hospital das Clínicas, São Paulo, Brazil
- Tokyo Medical AI Lab, Japan
- Red Cross Children's Hospital, Cape Town, South Africa
Global Ethics Council ensures privacy safeguards and algorithmic transparency.
Future Directions
Next-phase improvements include federated learning to remove central data pooling, expanded cohort enrollment from South Asia, and patient-facing explainability dashboards.
- • Collaboration with WHO Rare Disease Working Group for standardization.
- • Open benchmark challenges inviting external validation.
- • Funding pipeline via Global Innovation Fund to support scale-up.