Improve Patient Risk Assessment with Custom AI

Urgent care centers often rely on the built-in alerting features of their EHR systems, like Experity or DocuTAP. These systems typically use a static, rule-based logic such as the Modified Early Warning Score (MEWS). MEWS is helpful but limited, calculating a score based on a handful of vital signs like heart rate and blood pressure. The system cannot interpret the rich, unstructured text in a triage nurse's notes.

Consider this scenario: a 70-year-old patient arrives with a cough. Their vitals are stable, so their MEWS score is a benign 1. The triage nurse, however, types "patient seems confused and lethargic" into the free-text notes field. The rule-based EHR alert system completely ignores this text. It sees only the normal vitals. Hours later, the patient's condition deteriorates rapidly from sepsis, an outcome the clinical note hinted at but the rigid scoring system missed.

The structural problem is that EHRs are designed as systems of record, not predictive analytical engines. Their architecture prioritizes data storage and retrieval over real-time, complex inference. Getting new logic into an EHR often involves long and expensive change requests with the vendor, and they are not equipped to handle custom machine learning models. You are stuck with one-size-fits-all rules that miss the critical nuances in your specific patient population.

The engagement would start with a data audit and signing a Business Associate Agreement (BAA) to ensure HIPAA compliance. Syntora would analyze 12-24 months of de-identified data from your EHR to identify the most predictive features for a target outcome, such as hospital admission or sepsis onset. This audit establishes if there is sufficient signal in your historical data to build a high-performing model.

The technical architecture would use the Claude API to parse unstructured triage notes and extract key clinical entities. These extracted features, combined with structured data like vitals and labs, would feed a gradient boosting model built with Python's LightGBM library. The entire application would be packaged as a FastAPI service and deployed on AWS Lambda, creating a secure, serverless, and HIPAA-compliant endpoint that scales automatically with demand and has a processing time under 500ms.

The delivered system connects to your EHR. When a patient is triaged, your EHR sends the patient's data to the secure API. The API returns a risk score from 1 to 100 and the top three contributing factors. This information is written directly into a custom field in the patient's chart, providing your clinical team with immediate, explainable, and actionable insight without needing to log into a separate dashboard.