Abstract P386: Reducing Laboratory Testing in Hospitalized Stroke Patients Using Machine Learning

Abstract

Introduction: Despite previous studies indicating a high rate of overuse, in-hospital laboratory blood testing remains common. Repeated laboratory tests often return normal and lead to increased cost and even anemia. We hypothesize this situation is highly significant among hospitalized stroke patients. We used machine learning to reduce the freqeuency of repeated lab testing for stroke patients as a proof of concept. Method: We used stroke patients in the MIMIC III database to develop a novel deep learning algorithm that takes into account the characteristics of all study patients to predict whether repeating a lab test is necessary and aid in decisions of obtaining subsequent laboratory tests. Our model jointly considers temporal and spatial correlations of lab tests to make predictions of future test abnormality. We included 12 tests for electrolytes, renal function, and Complete Blood Count. We defined an error as recommending the reduction of a laboratory study when the testing would have yielded a critical abnormal result which should have been tested (Table 1). Results: We analyzed 3,178 stroke patients (293,433 lab results). Our training cohort was used to develop the algorithm and the testing data were used to generate results. In the testing cohort, we could maintain a 99.8% accuracy with a 52% reduction of all lab tests, which corresponded to a total reduction of 29,803 lab tests (74 false negatives). Different labs can be reduced at different rates. K and Cl were relatively easy to predict a critical abnormality on repeat testing while Ca and Mg displayed boundaries that were harder to predict abnormalties. Conclusions: Reducing lab testing by 52% from this stroke patient cohort maintained accuracy of testing at 99.8%. Our data suggest that multiple or daily laboratory orders are unlikely to change patient management given the high accuracy proposed by the machine learning approach. Table 1: normal range and critical abnormality for the 12 tests of interest