Predicting Anabolic Androgenic Steroid Doping among Specialized Health Care Patients with Elastic Net Regression Reveals Potential Laboratory Variables for “Patient Biological Passport”

TL;DR

This study explores using lab data to detect non-medical steroid use in patients, finding that certain lab variables can predict doping with reasonable accuracy.

Contribution

The study introduces a method to detect anabolic androgenic steroid use in clinical patients using elastic net regression and longitudinal lab data.

Findings

Elastic net regression models showed promising ability to detect AAS doping in clinical patients.

Using out-of-reference range lab measurements as predictors achieved an AUC of 0.757.

The study suggests real-life lab data can be used to detect AAS doping in a clinical context.

Abstract

Recent years have brought significant development in athlete doping use detection with the implementation of the Athlete Biological Passport (ABP). The aim of this study was to explore if similar methods could also be used to detect non-medical use of anabolic androgenic steroids (AAS) among clinical patients. For this purpose, six elastic net regression models were trained in a sample of Finnish specialized health care male patients (N = 2918; no doping = 1911, AAS doping = 1007), using different approaches to longitudinal laboratory measurements as predictive variables. The laboratory data was retrieved from the Hospital District of Helsinki and Uusimaa (HUS) data lake, and doping use status was defined by patient disclosure, recorded in digital medical record free texts. Length of observation time (e.g., time between the first and last laboratory measurement) was used as weight. Model performance was tested with holdout cross-validation.

All the tested models showed promising discriminative ability. The best fit was achieved by using the existence of out-of-reference range measurements of 31 laboratory parameters as predictors of AAS doping, with test data area under the receiver operating characteristic curve (AUC) of 0.757 (95% CI 0.725–0.789).

The findings of this preliminary study suggest that AAS doping could be detected in clinical context using real-life longitudinal laboratory data. Further model development is encouraged, with added dimensions regarding the use of different AAS substances, length of doping use, and other background data that may further increase the diagnostic accuracy of these models.

The online version contains supplementary material available at 10.1186/s40798-025-00854-5.

Figures