An artificial intelligence model integrating clinico-laboratory data and single nucleotide polymorphism-based genomic risk for prostate cancer diagnosis in Korean men.
Prostate cancer (PCa) is traditionally diagnosed using prostate-specific antigen (PSA)-based testing together with demographic and clinical factors. Building on this framework, we aimed to develop an AI (artificial intelligence) model for prebiopsy PCa diagnosis by integrating Korean population-relevant risk-associated single nucleotide polymorphisms (SNPs) to improve diagnostic accuracy.
Three models were developed in this study: Korean PCa-specific genomic score (GenPCa-Kor score), electronic medical record (EMR) meta-model, and Geno-EMR meta-model. From genome-wide association study summary statistics, 1,347 PCa-associated SNPs were selected for a deep neural network to derive the GenPCa-Kor score. Thirteen clinico-laboratory EMR parameters were used to build a stacking ensemble (EMR meta-model) with Light Gradient Boosting Machine, and Histogram-based Gradient Boosting Machine, and logistic regression as base learners and logistic regression as the meta-learner, using 10-fold cross-validation and Bayesian hyperparameter optimization. The Geno-EMR meta-model added the GenPCa-Kor score as a 14th feature to the same architecture.
Of 1,590 systematic biopsy-confirmed participants, 1,006 were analyzed; 757 comprised the training cohort and 249 consecutive patients comprised the independent test cohort. In the training cohort, the EMR meta-model and Geno-EMR meta-model achieved area under curves (AUCs) of 0.868 and 0.924, respectively. In the test cohort, their AUCs were 0.859 and 0.892, respectively. For clinically significant PCa (Grade Group ≥2), the Geno-EMR meta-model further improved the AUC from 0.887 to 0.911.
The Geno-EMR meta-model that integrate routine clinico-laboratory parameters with the SNP-based GenPCa-Kor score showed improved discrimination for PCa compared with the EMR meta-model alone.
Three models were developed in this study: Korean PCa-specific genomic score (GenPCa-Kor score), electronic medical record (EMR) meta-model, and Geno-EMR meta-model. From genome-wide association study summary statistics, 1,347 PCa-associated SNPs were selected for a deep neural network to derive the GenPCa-Kor score. Thirteen clinico-laboratory EMR parameters were used to build a stacking ensemble (EMR meta-model) with Light Gradient Boosting Machine, and Histogram-based Gradient Boosting Machine, and logistic regression as base learners and logistic regression as the meta-learner, using 10-fold cross-validation and Bayesian hyperparameter optimization. The Geno-EMR meta-model added the GenPCa-Kor score as a 14th feature to the same architecture.
Of 1,590 systematic biopsy-confirmed participants, 1,006 were analyzed; 757 comprised the training cohort and 249 consecutive patients comprised the independent test cohort. In the training cohort, the EMR meta-model and Geno-EMR meta-model achieved area under curves (AUCs) of 0.868 and 0.924, respectively. In the test cohort, their AUCs were 0.859 and 0.892, respectively. For clinically significant PCa (Grade Group ≥2), the Geno-EMR meta-model further improved the AUC from 0.887 to 0.911.
The Geno-EMR meta-model that integrate routine clinico-laboratory parameters with the SNP-based GenPCa-Kor score showed improved discrimination for PCa compared with the EMR meta-model alone.
Authors
Jung Jung, Han Han, So So, Hong Hong, Ahn Ahn, Lim Lim, Han Han, Park Park, Lee Lee, Song Song, Choi Choi, Chung Chung, Chung Chung, Kang Kang, Eom Eom, Koh Koh, Kim Kim
View on Pubmed