METHODOLOGICAL CHARACTERISTICS OF PHARMACOGENOMICS-INTEGRATED AI CLINICAL DECISION SUPPORT SYSTEMS (AI-CDSS): A SYSTEMATIC REVIEW
Author(s)
Shinyoung Park, PharmD1, Hae Sun Suh, MA, MS, PhD2;
1Kyung Hee University, Department of Regulatory Science, Graduate School, Seoul, Korea, Republic of, 2Kyung Hee University, College of Pharmacy, Seoul, Korea, Republic of
1Kyung Hee University, Department of Regulatory Science, Graduate School, Seoul, Korea, Republic of, 2Kyung Hee University, College of Pharmacy, Seoul, Korea, Republic of
OBJECTIVES: Pharmacogenomics (PGx) has the potential to support patient-centered precision medicine; however, evidence on artificial intelligence-based clinical decision support systems (AI-CDSS) integrating PGx remains fragmented. This systematic review examined the methodological characteristics of PGx-integrated AI-CDSS modeling and validation studies.
METHODS: Studies published through January 1, 2026 were identified in PubMed, EMBASE, the Cochrane Library, and the ACM Digital Library following PRISMA guidelines. Search terms covered AI or machine learning, clinical decision support systems, pharmacogenomics, and modeling or validation. We included studies reporting AI-CDSS model development or validation incorporating PGx data. Risk of bias and applicability were assessed using PROBAST. Extracted information included clinical domain, PGx inputs, analytical methods, validation strategies, explainability techniques, and performance metrics.
RESULTS: After duplicate removal, 147 unique records were screened, and 18 studies met the inclusion criteria; eight were assessed as having low risk of bias and adequate applicability. Ten studies were assessed to have high risk of bias, primarily due to small sample sizes along with model overfitting concerns. More than half focused-on oncology or individualized treatment optimization, such as warfarin dose prediction or ovarian stimulation in in vitro fertilization. Across studies, PGx-integrated AI-CDSS models combined genomic variants with structured clinical and biomarker data to support patient-level decision-making. Frequently applied tree-based ensemble methods, particularly XGBoost and Random Forest, generally outperformed traditional models. Internal validation using 5- or 10-fold cross-validation was commonly used. Explainability methods, including SHAP and LIME, were used to interpret the contribution of genetic and clinical features. Model performance was primarily reported using AUC-ROC (range: 0.62-0.97), while regression and alternative classification outcomes were evaluated using mean absolute error and F1-score.
CONCLUSIONS: PGx-integrated AI-CDSS studies showed methodological innovation; however, a substantial proportion are limited by small sample sizes and associated risks of overfitting. Further research using larger datasets and rigorous validation is needed for safe clinical implementation.
METHODS: Studies published through January 1, 2026 were identified in PubMed, EMBASE, the Cochrane Library, and the ACM Digital Library following PRISMA guidelines. Search terms covered AI or machine learning, clinical decision support systems, pharmacogenomics, and modeling or validation. We included studies reporting AI-CDSS model development or validation incorporating PGx data. Risk of bias and applicability were assessed using PROBAST. Extracted information included clinical domain, PGx inputs, analytical methods, validation strategies, explainability techniques, and performance metrics.
RESULTS: After duplicate removal, 147 unique records were screened, and 18 studies met the inclusion criteria; eight were assessed as having low risk of bias and adequate applicability. Ten studies were assessed to have high risk of bias, primarily due to small sample sizes along with model overfitting concerns. More than half focused-on oncology or individualized treatment optimization, such as warfarin dose prediction or ovarian stimulation in in vitro fertilization. Across studies, PGx-integrated AI-CDSS models combined genomic variants with structured clinical and biomarker data to support patient-level decision-making. Frequently applied tree-based ensemble methods, particularly XGBoost and Random Forest, generally outperformed traditional models. Internal validation using 5- or 10-fold cross-validation was commonly used. Explainability methods, including SHAP and LIME, were used to interpret the contribution of genetic and clinical features. Model performance was primarily reported using AUC-ROC (range: 0.62-0.97), while regression and alternative classification outcomes were evaluated using mean absolute error and F1-score.
CONCLUSIONS: PGx-integrated AI-CDSS studies showed methodological innovation; however, a substantial proportion are limited by small sample sizes and associated risks of overfitting. Further research using larger datasets and rigorous validation is needed for safe clinical implementation.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
MSR244
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
No Additional Disease & Conditions/Specialized Treatment Areas