利用非侵入性臨床參數建構雙重AI預測模型以協助
男性膀胱出口阻塞的診斷並減少錄影尿動力學檢查的需求

蔡宗佑^1,2*、田靜慧³、李峻嘉¹、郭漢崇⁴

¹亞東紀念醫院外科部泌尿外科, ²元智大學電機工程學系, ³花蓮慈濟醫院研究部, ⁴花蓮慈濟醫院泌尿部

Developing Dual AI Models Using Non-Invasive Clinical Parameters to Aid in Male Bladder Outlet Obstruction Diagnosis and Minimize the Need for Invasive Video-Urodynamic Studies

Chung-You Tsai^1,2*, Jing-Hui Tian³, Jiun-Jia Li¹, Hann-Chorng Kuo⁴

¹ Divisions of Urology, Department of Surgery, Far Eastern Memorial Hospital, Taiwan.
² Department of Electrical Engineering, Yuan Ze University, Taiwan. ³ Department of Medical Research, Hualien Tzu Chi Hospital, Taiwan. ⁴ Department of Urology, Hualien Tzu Chi Hospital, Buddhist Tzu Chi Medical Foundation and Tzu Chi University, Taiwan

Purpose: Video-urodynamic studies (VUDS) can discern among male bladder outlet obstruction (BOO) causes such as PBNO, BPO, US, and DV, potentially outperforming pressure-flow studies. Despite this, its invasive nature prompts this study to create non-invasive predictive models, aimed at aiding diagnosis and reducing the need for invasive examinations.

Materials and Methods: A retrospective analysis of male patients experiencing lower urinary tract symptoms (LUTS) who underwent VUDS between January 2001 and May 2022 in a single medical center was performed. Two predictive models, based on binary classification criteria for the presence or absence of obstruction, namely International Continence Society-defined BOO (ICS_BOO), and VUDS-diagnosed BOO (VBOO), were developed. The dataset, consisting of 307 patients, was divided randomly into a training set (70%, n=215) and a test set (30%, n=92). Six machine learning algorithms (LR, SVM, DT, RF, GBDT, and XGBoost) were used in model construction, with 20 iterations of 5-fold cross-validation performed for internal validation. The best-performing model was chosen for external validation with the test set.

Results: Of the 307 patients, 82 (26.7%) met ICS_BOO criteria, 255 (73.3%) were non-ICS_BOO, 252 (82.1%) were diagnosed as VBOO, and 55 (17.9%) as non-VBOO. The LR-based nomogram for ICS_BOO prediction achieved an AUC of 0.74±0.09 (internal validation, accuracy 0.76±0.04) and 0.86 (external validation, accuracy 0.77). For VBOO prediction, it achieved an AUC of 0.71±0.06 (internal validation, accuracy 0.77±0.06) and 0.72 (external validation, accuracy 0.76). If dual-model predictions showed both ICS_BOO and VBOO as positive in external validation, all patients (100%) were either pure BPO or PBNO, suggesting that VUDS may be avoided for 30 patients (32.6%) with dual-model positive predictions.

Conclusion: Two machine learning models predicting ICS_BOO and VBOO, based on non-invasive clinical parameters, demonstrate commendable discrimination performance. When both models predict positively, VUDS may be avoided, assisting in male BOO diagnosis and potentially reducing the need for such invasive procedures in approximately one-third of the patients.