Development Of A Machine Learning Polyvariant Risk Prediction Model For Severe Cutaneous Adverse Drug Reactions To Carbamazepine And Other Aromatic Antiseizure Medications

Importance. Carbamazepine (CBZ) is a first-line treatment for epilepsy and neuropathic pain but carries risk for severe and sometimes life-threatening cutaneous reactions. Across Asian populations, the human leukocyte antigen HLA-B*15:02 in chromosome 6 is strongly associated with these adverse reactions, but the estimated positive predictive value (PPV) is only 0.07.

Objective. To develop a polyvariant risk prediction model with improved PPV.

Design, Settings and Participants. Case-control study. Cases had severe cutaneous reactions (Stevens-Johnson syndrome [SJS] or toxic epidermal necrolysis [TEN]) induced by carbamazepine (n=84) or other aromatic antiseizure medications (n=28). Cases were matched with drug-tolerant controls (n=80). All cases and controls were of Han Chinese descent, and 56% (107/192) carried HLA-B*15:02.

Exposures. Carbamazepine and other aromatic antiseizure medications.

Main Outcomes and Measures. All participants underwent whole genome sequencing. Putative genetic variants that modify the risk of developing carbamazepine-induced SJS/TEN were identified using a machine-learning workflow. Different model types were evaluated based on the weighted F1 (WTD F1) and area under the receiver operating characteristic curve (AUC) scores, and using genotypes within the known risk region (chromosome 6). The best-performing model was then used to screen for additional putative genetic markers. The bootstrapped PPV value obtained from the resulting polyvariant model was compared to that estimated from HLA-B*15:02 carrier status.

Results. We identified 465 common variants that collectively achieved a median AUC of 0.98 (95% confidence interval [CI]: 0.9799-0.9828) in predicing carbamazepine-induced SJS/TEN. Bootstrap analysis showed the model had a median PPV of 0.20 (95% CI: 0.19-0.22) and a negative predictive value of 1.0. In cases of SJS/TEN induced by aromatic medications other than CBZ, the model achieved AUC of 0.82 (95% CI: 0.82-0.83) and WTD F1 of 0.76 (95% CI: 0.76-0.77), indicating good generalisability.

Conclusion and Relevance. A machine learning approach is able to enhance the genetic prediction of severe cutaneous reactions induced by aromatic antiseizure medications. Improved precision of genetic risk prediction could stratify HLA-B*15:02 carriers to receive CBZ treatment with increased safety.

Recommended citation: Zeibich, Robert; Anderson, Alison; Chen, Zhibin; Shi, Yi-Wu; Ng, Ching-Ching; Baum, Larry; Cherny, Stacey; Sham, Pak-Chung; Lim, Kheng Seang; Liao, Wei-Ping; O’Brien, Terence J.; Perucca, Piero; and Kwan, Patrick. Development of a Machine Learning Polyvariant Risk Prediction Model for Severe Cutaneous Adverse Drug Reactions to Carbamazepine and Other Aromatic Antiseizure Medications. Preprint: THE LANCET. 2024 Nov 9. View article online
Download Paper