Computer Assisted Detection of Rare Kidney Diseases in Emergency Departments

Details

Supervisors

Devuyst, Olivier ; Gillion, Valentine

Faculty

Faculté de médecine et médecine dentaire

Degree label

Master [180] en médecine, à finalité spécialisée

Abstract

Importance. There are at least 500 rare kidney diseases presenting with variable manifestations, making diagnosis long and challenging even in specialized centers. Patients carrying rare disorders often spend years before getting a diagnosis, leaving the disease uncontrolled and worsening the outcomes. Objectives. To assess how the introduction of electronic health records could accelerate the recognition of patients with rare kidney diseases by developing machine learning models to assist with their detection. Design, Setting and Participants. Diagnostic study of a nationally representative sample of 290 905 Emergency Department (ED) visits in the United States of America from the NHAMCS from January 1, 2009, through December 31, 2020. The analysis was conducted in 2022. Main Outcome. The principal outcome was the diagnosis of a rare kidney disease during the visit or hospitalisation following the visit. We split the data into a training set composed of 80% of the negative samples and trained a deep neural network auto-encoder. We then derived two additional models by randomly selecting 25% of the positive cases and comparing them with the remaining test cases using cosine similarity and then combining the two models. We measured the predictive performance by generating the commonly used predictors sensitivity, specificity, Positive Likelihood Ratio (PLR), Negative Likelihood Ratio (NLR), Positive Predictive Value (PPV), Negative Predictive Value (NPV) and F-Score. Results. In a sample of 290 905 ED visits, 20 were considered positive (0.0069%). The three models designed generated statistically significant results when comparing the score distributions of negative and positive cases (p < 10−6). Conclusions. Machine learning models can assist with the detection of rare kidney diseases using readily available data in Electronic Health Records (EHR) both in primary and secondary care.