Predicting laboratory tests from patient generated health data could accelerate laboratory orders in the emergency department

Overcrowding and diagnostic delays compromise patient care in the emergency department. This study investigates the use of machine learning to preemptively predict a broad set of laboratory test orders (CRP, Ddimer, GOT, NT-proBNP, PCT, hs-TnT) solely from patientreported data, collected via a digital symptom checker at a tertiary emergency department. Using random forests on our dataset consisting of 3925 patients, we found ROC-AUCs ranging from 0.74 (CRP) to 0.88 (GOT). The time from symptom check to test order (up to 6.5 hours) highlights the potential for saving considerable time. Early ROC curves showed that frequent tests like GOT performed well enough for preemptive ordering, while low-frequency, high-stakes tests, such as D-dimer and PCT may be better utilized as prompts for clinical decision-making. These findings show that data collected at triage time gained through a structured patient-reported questionnaire can offer a way to reduce the time-to-result and enhance ED efficiency.