Related Experiment Video
Updated: Jul 2, 2026
Using Reference Reagents to Confirm Robustness of Cytokine Release Assays for the Prediction of Monoclonal Antibody Safety
Published on: September 15, 2023
Multi-center validation of a machine learning model for early detection of monoclonal immunoglobulin-related
Yujiao Hu1, Xiaoyan Hao1, Xiaoyan Li1
1Department of Clinical Laboratory Medicine, Xijing Hospital, Fourth Military Medical University, Xi'an, China.
Background:
Monoclonal immunoglobulin (MIg)-related disorders are clonal plasma cell diseases defined by aberrant MIg overproduction, typically presenting with subtle, non-specific manifestations that lead to underrecognition in routine clinical practice, delayed diagnosis, and subsequent irreversible organ damage and poor outcomes.
Methods:
To address this critical gap and facilitate timely MIg diagnosis, we developed and compared eight machine learning models using clinical and routine laboratory data. Cohort 1 (n=5018) included patients undergoing their first serum immunofixation electrophoresis (IFE) test at Xijing Hospital between 2017 and 2024, comprising 2038 MIg-positive and 2980 MIg-negative cases. The top-performing model was subsequently advanced for Cohort 2 (n=2270), consisting of three subsets (Set 1: Xijing Hospital, n=1753; Set 2: 3201 Hospital, n=398; Set 3: Shaanxi Provincial People's Hospital, n=119), served as a fully independent multi-institutional cohort, and was applied to assess the generalizability of the model.
Results:
Feature selection identified ten key variables-urinary cast count, mean corpuscular volume, prothrombin time, activated partial thromboplastin time, albumin/globulin ratio, globulin, total protein, calcium, age, and sex-which were used for final model construction. Among the eight machine learning algorithms, the light gradient boosting machine (LightGBM) model showed superior performance, achieving an AUC of 0.945 (87.0% sensitivity, 89.6% specificity) in the Cohort 2, and achieved 84.6% sensitivity, significantly outperforming the conventional serum protein electrophoresis (75.3% sensitivity) in the subset of cohort 2. In a separate cohort of newly enrolled patients, its 88.2% sensitivity also exceeded that of three non-specialist clinicians.
Conclusions:
This LightGBM model, built on readily available clinical and routine laboratory data, may help identify potential MIg-related cases and prioritize patients for further specialized confirmatory testing, which could facilitate earlier recognition and reduce diagnostic delays.