Development of Digital Algorithms for the Early Identification of Rare Clinical Outcomes Using Electronic Medical Records: A Case Study of Hidradenitis Suppurativa

Abstract

Hidradenitis suppurativa (HS) is a painful, progressive, and disfiguring rare inflammatory skin condition with significant diagnostic delays due to limited awareness. This study aims to develop and validate digital algorithms using integrated electronic medical records (ieMR) to identify undiagnosed HS patients. A test cohort of 121 HS cases (patients who visited a dermatologist-led HS clinic and received treatment) and 187,106 controls (emergency department patients) were identified from patients attending a Queensland tertiary hospital from 2018 to 2022. Using demographics, structured ieMR data and free-text flags, we developed one logistic regression and two random forest algorithms (with and without class weighting) to predict HS. A clinical chart audit of 200 randomly selected patients helped refine the best-performing algorithm, which was then validated using data from another tertiary hospital. Logistic regression performed best at a threshold of 0.4 (sensitivity: 0.66 [0.58, 0.74], positive predictive value [PPV]: 0.71 [0.64, 0.80]). Strong predictors included dermatology clinic visits, free-text diagnostic notes, lesion location terms, antibiotic and isotretinoin use, and elevated inflammatory markers. Internal validation showed high agreement, and the refined model improved sensitivity to 0.89 (0.83, 0.94) and PPV to 0.87 (0.81, 0.92). This model performed relatively well in the validation cohort, with sensitivity and specificity (at threshold 0.4) both >0.70 and a PPV of 0.45, supporting clinical utility. Validated digital algorithms incorporating key diagnostic indicators may help identify undiagnosed HS patients, reducing diagnostic delays and improving prevalence assessment.