Using Machine Learning and Contextual Information to Decrease the Time Needed to Identify Postoperative Complications Retrospectively

To improve the quality of surgical care, it seems to be increasingly important to assess and compare the delivery of healthcare across institutions. A key aspect of this is the monitoring of postoperative complications following surgical procedures.
However, the process of identifying risk factors and complications, often recorded as unstructured data in Electronic Health Records (EHRs), is costly and time-consuming. To address this, we developed a machine learning (ML) model that utilizes structured data from patients' EHRs to identify those with one or more postoperative complications, achieving a correct classification rate of 79%.
Furthermore, we extracted summaries of structured data (contextual information) from the EHRs for each patient, offering clues about potential complications.
Contextual Information as well as probabilities for complications provided beforehand should enable the data nurses to search more focused for hints of complications in the EHR data. This approach aims to enhance the efficiency and accuracy of identifying postoperative complications by study nurses, thereby improving patient care, and reducing the workload on clinical staff.

The aim of this study was to determine whether employing ML models to accurately predict complications and offering contextual information in advance could decrease 1) the time of data collection and 2) the time until the identification of a complication (time-to-complication).

The results show that using ML models and contextual information leads to a significant reduction in data collection time and a faster identification of a complication. The average collection time in the intervention group was 14.22 minutes compared to 16.53 minutes in the control group, which is a significant difference (p<0.043).

Additionally, time-to-complication was significantly shortened in the intervention group (p<0.001). The ML models used showed high accuracy in identifying actual complicative cases (82.76%).

A randomized, controlled study, conducted from August to September 2023, included the collection of 391 cases. The methodology involved the implementation of ML models for the retrospective identification of complication cases, including assessments of the suspected type of complication (e.g. wound infection, UTI etc.) . Additionally, context information was provided, which included structured data on the patients, such as primary and secondary diagnoses as well as specific laboratory or radiologic services characteristic for the diagnosis of a specific type of complication during the main stay. This was generated from EHR data and provided for the data nurses before the start of the analysis for each individual patient.

Thus, we were able to demonstrate that the information provided to the study nurses in advance (application of ML Models, contextual information) led to a significant reduction of 15% in the data collection process.