Computer-aided, Evidence-based System Improved Clinical Diagnostic Accuracy of Certificated-Physicians in Acute Abdominal Pain (Preprint)

Abstract

Acute abdominal pain (AAP) is a common complaint and can be caused by a broad spectrum of diseases. Mis-diagnosis or delay can lead to severe complications and increased mortality. The diagnostic performance of AAP remains sub-optimal.

We aimed to develop an artificial intelligence (AI) system for AAP diagnosis, and to evaluate the efficacy of the system.

The system consisted of 164 feature extraction models and 1 disease prediction model. The system will first match the extracted features with the pretherapeutic/preoperative diagnostic criteria for each AAP disease and output the corresponding diagnose. If none of the criteria was matched, the system will make a feature-based fitting for clinical diagnosis prediction. Pre-training, training and validation of the system included 65,933, 2766 and 607 AAP patients, respectively. A randomized control reader study was conducted including 22 physicians, 11 in the AI-assist group and 11 in the control group to further evaluate the performance of system.

The average accuracy for feature extraction was 97.52%. For disease prediction, the system achieved accuracy of 84.18% for clinical diagnosis. In the reader study, both accuracies of clinical diagnosis and final diagnosis were significantly higher in AI-assisted group compared to control group (81.40% vs 69.49%, P＜0.001, OR=0.511, 95% CI. = 0.225-0.797; 92.54% vs 86.27%, P＜0.001, OR=0.518, 95% CI. = 0.133-0.258, respectively).

The system achieved an excellent performance in predicting AAP diseases and has great potential to become a clinically useful diagnostic aid in AAP diagnosis.