Machine learning algorithms that integrate multiple biomarkers are increasingly used in disease detection, yet economic considerations are often overlooked. Medial vascular calcification (mVC), a pathology associated with elevated cardiovascular risk in chronic kidney disease (CKD), requires cost-effective diagnostic approaches. This pilot study evaluated the cost-effectiveness of machine learning models for mVC detection using traditional risk markers and circulating biomarkers in 152 CKD patients undergoing living donor kidney transplantation. Patients were classified as having no/minimal (n = 93) or moderate/extensive (n = 59) mVC. Five classification frameworks with automatic variable selection identified predictors of mVC. Age and copeptin were selected by all algorithms, while diabetes, male sex, choline, and osteoprotegerin were chosen by four methods. The number of features selected ranged from 5 to 21. Although accuracy differences among classifiers were limited to 3%, models using more features nearly tripled the procedure's cost. By incorporating the incremental cost-effectiveness ratio, the study highlighted significant disparities in performance versus cost among classifiers. The present findings suggest that machine learning has the potential to complement imaging techniques for mVC detection and uncover novel biomarkers. However, modest performance improvements may not justify higher costs, underscoring the importance of considering cost-effectiveness when selecting classification models.
Background: Evidence to help avoid unnecessary prostate biopsies is being actively pursued. Our goal was to develop and internally validate a nomogram for predicting clinically significant prostate cancer (csPC) in men with low suspicion of disease (prostate specific antigen [PSA] < 10 ng/mL, normal digital rectal examination [DRE]), in whom magnetic resonance imaging (MRI) findings are positive. Methods: Patients with no prior prostate cancer diagnosis who underwent MRI–ultrasound fusion biopsy of the prostate were retrospectively analyzed. Inclusion criteria were PSA < 10 ng/mL, normal DRE, Prostate Imaging Reporting And Data System (PIRADS) category ≥ 3, and no extraprostatic extension or seminal vesicle invasion reported on MRI. Associations between csPC diagnosis and patient or lesion characteristics were analyzed, and a multivariable model was developed. Internal validation of the model with 5-fold cross-validation and bootstrapping methods was performed. Results: Among 209 patients, 67 were diagnosed with csPC. Factors incorporated into the model for predicting csPC were age, 5-alpha reductase inhibitor use, PSA, prostate volume, PIRADS > 3, and lesion location in the peripheral zone. The model’s ROC AUC was 0.86, with consistent performance at internal validation (0.84 with cross-validation, 0.82 with bootstrapping). With an empirical threshold of <10% csPC probability to omit biopsy, 72 (50.7%) unnecessary biopsies would have been avoided, at the cost of missing 2 (3.0%) csPC cases. Conclusions: Our nomogram might serve as a valuable tool in refining selection criteria in men considered for prostate biopsy. The major limitation of the study is its retrospective character. Prospective, external validation of the model is warranted.