¹National Institute of Mental Health, National Institute of Neurology and Psychiatry, Department of Neuropsychopharmacology, Tokyo, Japan, ²Tokyo Metropolitan Institute of Medical Science, Department of Clinical Medical Sciences, Tokyo, Japan, ³Saitama Medical University International Medical Center, Department of Anesthesiology, Saitama, Japan, ⁴Juntendo University School of Medicine, Department of Anesthesiology and Pain Medicine, Tokyo, Japan, ⁵Tokyo Dental College, Department of Oral Health and Clinical Science, Tokyo, Japan

Background: Opioid analgesics are widely used for pain control, but individual responses vary greatly, partly due to genetic differences. To investigate individual differences, Yoshida et al. (PLOS ONE, 2015) used multiple linear regression to predict 24-hour postoperative fentanyl use from single-nucleotide polymorphisms (SNPs) and clinical variables, reporting R² values of 0.15 (p = 5.66 x 10^-10) and 0.033 (p = 0.030) on two independent datasets. Yoshida et al. findings highlighted the potential utility of genetic markers in guiding opioid dosing but also revealed limitations in model generalizability.

Aims & Objectives: To explore whether a machine learning approach can enhance generalizability and interpretability, we applied support vector machine (SVM) regression to the dataset originally analyzed by Yoshida et al. (PLOS ONE, 2015).

Method: The study protocol was approved by the Institutional Review Boards of Tokyo Dental College, the Institute of Medical Science at the University of Tokyo, Toho University Sakura Medical Center, and the Tokyo Institute of Psychiatry (now the Tokyo Metropolitan Institute of Medical Science). Written informed consent was obtained from all participants, including parental consent for minors.
Here, we used two datasets, originally described in Yoshida et al, retrospectively. The first dataset (n = 354) comprised healthy patients (15–52 years, 126 males, 228 females) undergoing cosmetic orthognathic surgery under general anesthesia. The second dataset (n = 145) included patients (28–80 years, 83 males, 62 females) undergoing major abdominal surgery under combined general and epidural anesthesia. Body weight and height were recorded, and peripheral blood or oral mucosa samples were collected for genetic analysis. Genotyping was conducted for five SNPs: OPRM1 rs9384179, CACNA1E rs3845446, ADRB2 rs11959113, GIRK2 rs2835859, and CREB1 rs2952768.
An SVM regression model was implemented in MATLAB 2023b to predict postoperative analgesic opioid use from body weight, age, sex, and the five SNPs. Model performance was evaluated using leave-one-out cross-validation (LOOCV) on the training dataset (n = 354). The trained model was then tested on the independent dataset (n = 145). Shapley Additive Explanations (SHAP) was used to evaluate feature importance.

Results: The results of LOOCV showed an R² of 0.12 (p = 1.52 × 10⁻¹¹), while testing on the external dataset yielded an R² of 0.040 (p = 0.011). SHAP analysis indicated that ADRB2 had the greatest importance among the evaluated SNPs.

Discussion & Conclusions: Compared to the original regression model by Yoshida et al., which reported R² values of 0.15 (training) and 0.033 (test), the SVM model demonstrated similar test performance but a slightly lower training R² of 0.12. These differences may be attributed to variations in feature selection and validation methods. Yoshida et al. excluded age, sex, and CACNA1E but included pain perception latency (PPL) as one of their predictors. Additionally, they trained their model on the full dataset without cross-validation, whereas we applied LOOCV to mitigate overfitting. Both studies indicated a relationship between SNPs and postoperative analgesic use; however, further validation is necessary to confirm these findings.