NHCS Cuts Lead Time for 6,128 Discharge Summaries from 286 Hours to 1 Hour with Pair

The National Heart Centre Singapore (NHCS) reduced the net lead time for 6,128 discharge summaries from 286 hours of manual interpretation to just 1 hour for prompt generation and upload in Pair, an efficiency gain of 99.6%. Pair is a fast and secure AI Assistant developed by Open Government Products (OGP). In an initial test on the discharge summaries of 600 patients, Pair achieved an overall accuracy of 96% across all 19 Charlson Comorbidity Index (CCI) components when compared with manual reviews performed by human auditors, and its errors were randomly distributed around a mean of zero. NHCS held a database of patient case notes multiple times larger than any other within the hospital, but IT policy restrictions made it challenging to develop in-house analytics for processing such a large volume of text data. The hospital chose to examine heart attack patients' increasing length of stay over the years. Associate Professor Anders Sahlen, Senior Consultant, Cardiology, and Director of Echocardiography Core Laboratory, and Lau Yee How, Senior Manager, Singapore Cardiac Data Bank, conducted a retrospective audit across thousands of patient case notes. The key first step was to extract the 19 components of the CCI, a general predictor of one-year mortality risk. Following the pilot, NHCS worked with the Pair team to increase Pair's processing ability to batch jobs of up to 5,000 case note entries simultaneously in a single document. Previously, each manual interpretation required 2.8 minutes of clinician time on average. Multiple Pair batch jobs can be executed simultaneously without human intervention, the entire batch processing was completed in 2.5 hours, and email notifications are sent upon completion. The prompt framework created for the use case can be shared with other teams within NHCS and SingHealth without the need for writing code or model training. Pair is still applied with a human in the loop to validate its output. The NHCS team has since moved on to apply Pair to other areas where extraction from unstructured text is required, such as clinic entries, radiology reports and patient feedback.