20 Years of Liver Biopsies at a Single Institution—Application of a Machine Learning‐Based Approach to Pathology Database Searches

ABSTRACT Advances in hepatology—such as noninvasive serum biomarkers (e.g., Fibrosure since 2004), imaging techniques (e.g., MR Elastography since 2009 and FibroScan since 2013), and effective direct‐acting antivirals for the treatment of chronic hepatitis C—have likely influenced liver biopsy practices. We assessed the impact of these developments on biopsy frequency and indications. Pathology database (2003–2022) was queried for reports containing ‘biopsy’ or ‘bx’ and ‘liver’ or ‘hepat’, identifying 32 842 cases. A manually reviewed subset of non‐consult, non‐intraoperative biopsies (22%) was categorised as: (1) lesional, (2) diagnostic (initial workup), (3) transplant (all allograft biopsies), or (4) staging/prognostic (in chronic liver disease). This manually classified dataset was divided into training and test sets to develop a machine learning algorithm evaluating the remaining 14 319 non‐classified cases. Autoregressive integrated moving average (ARIMA) model was applied to correlate changes in biopsy indications to interventions. Overall, the number of liver biopsies remained steady over 20 years, with declines in in‐house biopsies offset by an uptrend in outside consults. The algorithm was evaluated on the test set yielding an F1 score of 0.95, then applied to the non‐classified dataset, detecting proportional decreases in staging and transplant biopsies with proportional increases in diagnostic and lesional biopsies. ARIMA analysis identified a decline in allograft biopsies associated with the approval of direct‐acting antiviral agents. A machine learning approach enabled accurate large‐scale classification of liver biopsies, revealing evolving biopsy practices over time. This method may be applied across disciplines to evaluate the changing practices.