Automated Wrist Pulse diagnosis of Pancreatitis via Autoregressive Discriminant models

Wrist pulse signals, commonly used in traditional oriental medicine, reflects important pathological changes in the body which may be utilized to characterize an individual’s health status. Being one of the ‘four pillars of clinical evaluation’, pulse diagnosis plays a critical role in traditional Chinese medicine (TCM), informing the physician about crucial information such as the state of balance of the body and the state of internal organs. The traditional method of examination is via palpation in which the practitioner uses fingertips to feel the radial pulse of the patient. It is thus highly subjective and depends heavily on the practitioner’s experience. With the aid of technology, modern measurements of pulse signal can be taken in a more objective manner. However, there is a lack of tools and standards for analysing and interpreting these computerized pulse signals. In particular, diagnosis of inflammation is challenging and current clinical approaches via other methods (such as blood test, urine test, and X-ray) are time-consuming and often inaccurate. This paper proposes an automated procedure for distinguishing patients with inflammation using their digitized wrist pulse signals, based on a two-stage time-series classification technique involving autoregressive models followed by common classification methods such as Linear Discriminant Analysis (AR-LDA) and Logistic regression (AR-LR). We focus on one of the major inflammation symptoms – pancreatitis, a condition that can potentially lead to fatal complications in severe cases. We work with wrist pulse signals captured from patients using a Doppler ultrasonic blood analyser. After pre-processing, features derived from fitted AR models were fed to train a LDA and a LR classifier. The effectiveness of our approach is demonstrated using a subset of the wrist pulse database from Chen et al. (2009), consisting of 100 healthy persons and 54 Pancreatitis patients. To evaluate the classification performance, the models are trained on 50% of the data while the remaining observations were reserved for testing. An overall accuracy of 83% and 91% was achieved from AR-LR and AR-LDA respectively, both with an area under curve of 0.88. AR-LDA achieved a higher sensitivity (81%) and specificity (96%) with a positive predictive value of 92%. These results showed that AR-LDA is a promising approach for classifying time-series wrist pulse signals. It provides a relatively low-cost, easy-to-implement, and reliable tool for modern computerized wrist pulse diagnosis.