Claude Found Eleven Medical Errors in One Family's Records

Short summary

An engineer built a personal medical-record system (Node.js + SQLite) and used Claude Opus to audit family health history, finding 11 concrete errors across prescriptions, tests, and diagnoses. The post argues LLMs outperform existing systems at cross-record verification with full context, not conversational diagnosis—a structural advantage rooted in how medical records are fragmented across specialist EMRs. Cost: ~$0.60/patient/review; the architecture pattern is replicable for compliance and verification in regulated domains.

• •Engineer built a relational-database medical-record system and used Claude Opus to find 11 real errors in family history (missed tests, drug interactions, mislabeled prescriptions)
• •LLMs excel at cross-record verification with complete context, not at conversational diagnosis—the architecture provides all information in one prompt
• •Cost baseline: $0.60 per patient review at Opus rates; the JSON API pattern is replicable for any verification-focused regulated domain

Generated with AI, which can make mistakes.