GPT-assisted radiomic modeling for predicting pathological complete response to neoadjuvant chemoimmunotherapy in head and neck squamous cell carcinoma.
Authors
Affiliations (10)
Affiliations (10)
- School of Biomedical Engineering, Southern Medical University, ., Guangzhou, 510515, China.
- Department of Oral and Maxillofacial Surgery, Sun Yat-Sen Memorial Hospital, No. 107 Yanjiang West Road, Guangzhou, Guangdong Province, Guangzhou, 510120, China.
- Stomatological Hospital of Haizhu District, No. 4, Jiangnan Avenue North, Haizhu District, Guangzhou City, Guangdong Province, Guangzhou, 510220, China.
- School of Biomedical Engineering, Southern Medical University, No.1023,South Shatai Road, BaiyunDistrict, Guangzhou, China, Guangzhou, Guangdong, 510515, China.
- School of Biomedical Engineering, Southern Medical University, No.1023,South Shatai Road, BaiyunDistrict, Guangzhou, China, Guangzhou, 510515, China.
- Department of Radiology, Sun Yat-Sen Memorial Hospital, No. 107 Yanjiang West Road, Guangzhou, Guangdong Province, Guangzhou, 510120, China.
- Department of Radiology, Sun Yat-sen University Cancer Center, 651 East Dongfeng Rd, Yuexiu district, Guanzhou, Guangdong province, Guangzhou, 510060, China.
- Department of Medical Imaging, Guangzhou Institute of Cancer Research, No. 78 Hengzhigang, Yuexiu District, Guangzhou, Guangdong Province, Guangzhou, None Selected, 510095, China.
- The Stomatology Department of The First Affiliated Hospital, Medical College of Shantou University, 22 Xinling Road, Shantou, Guangdong, P.R.China, Shantou, 515041, China.
- School of Biomedical Engineering, Southern Medical University, No.1023,South Shatai Road, BaiyunDistrict, Guangzhou, China, Guangzhou, None Selected, 510515, China.
Abstract
To evaluate whether GPT can assist in automating radiomics modeling for predicting pathological complete response (pCR) to neoadjuvant chemoimmunotherapy in head and neck squamous cell carcinoma. The study included a training cohort (TC; n = 186), a validation cohort (VC; n = 116), and a prospective multicenter validation cohort (PVC; n = 269). Radiomic and supervised deep learning features were extracted from pretreatment T2-weighted MRI. Logistic regression (LR) was first evaluated using radiomic, deep learning, and fused features. Support vector machine, naïve Bayes, random forest, and XGBoost models were subsequently compared using fused features. GPT was not used as the classifier but assisted with preprocessing, feature selection, hyperparameter optimization, code generation and execution, model training, and probability output. Each GPT-assisted workflow was independently repeated five times. Fused features achieved the highest AUC for both manually developed and GPT-assisted LR models. Their AUCs were 0.759 versus 0.763 ± 0.003 in the TC, 0.714 versus 0.741 ± 0.008 in the VC, and 0.700 versus 0.706 ± 0.004 in the PVC. Across five classifiers, GPT-assisted workflows achieved performance broadly comparable to manually developed models. Repeated GPT-assisted runs showed limited variability, with AUC standard deviations ranging from 0.001 to 0.019. Under expert supervision, GPT can automate key radiomics modeling procedures while achieving performance comparable to conventional machine learning workflows, potentially reducing technical workload and facilitating integrated model development.