The Challenge
CareSync Health had partnered with three hospitals to pilot a remote patient monitoring programme for post-discharge cardiac and diabetic patients. The concept was sound — patients wore FDA-cleared biosensors at home that streamed vitals (heart rate, blood oxygen, blood glucose) to CareSync's servers. The problem was that no one could actually see and act on this data in a useful way. Clinicians were expected to log into a raw data dashboard that showed every reading from every patient as an undifferentiated stream of numbers. There was no alerting, no prioritisation and no clinical context. Nurses reported spending up to three hours per shift manually reviewing vitals data — and still missing critical changes because the volume made pattern recognition impossible. In the first month of the pilot, two patients who had deteriorating readings visible in the data were readmitted to hospital, with the data reviewed only after the fact.
Our Approach
We built a three-layer system: a HIPAA-compliant data ingestion and storage layer on AWS HIPAA-eligible services, an AI anomaly detection layer that flagged readings requiring clinical review, and a clinician-facing triage interface that presented patients in priority order with full vitals history and contextual alerts. The anomaly detection model was trained on a combination of published clinical thresholds (from the American Heart Association and ADA guidelines) and the historical data from CareSync's first month of readings. Rather than replacing clinical judgement, the model's job was narrow and specific: reduce the triage burden from reviewing every patient to reviewing the patients most likely to need intervention, ranked by urgency.
How We Built It
Clinical workflow research and HIPAA architecture (weeks 1–3)
We worked with CareSync's clinical consultant and a cardiologist from one of the pilot hospitals to understand triage workflow before designing any UI. Simultaneously, we designed the AWS architecture against the HIPAA Security Rule technical safeguards — encryption at rest (AES-256), encryption in transit (TLS 1.3), audit logging on all PHI access and strict IAM role boundaries.
Ingestion pipeline and storage layer (weeks 3–8)
Built the HL7 FHIR-compliant ingestion pipeline to receive vitals from three different biosensor manufacturers, normalise them to a common schema and store them in a partitioned PostgreSQL database on AWS RDS (HIPAA-eligible). Load tested to 10,000 concurrent device connections — 20× the current deployment size.
AI anomaly detection service (weeks 8–14)
Trained a gradient boosting model on 4.2 million historical readings to classify each new reading as normal, watch or alert. Built a human-in-the-loop review interface so clinical staff could mark false positives, which fed back into weekly model retraining. After six weeks of live operation, false positive rate dropped from 12% to 3.1%.
Clinician triage UI and hospital rollout (weeks 14–22)
Built the clinician dashboard with patient priority queue, 30-day vitals sparklines, alert history and one-click escalation to the on-call physician. Rolled out to three hospitals with dedicated training sessions. Gathered feedback from 24 nurses and 8 physicians across the pilot before the final production release.
