Regression analysis of radial artery pulse palpation as a potential tool for traditional Chinese medicine training education

Pulse palpation was an important part of the traditional Chinese medicine (TCM) vascular examination. It is challenging for new physicians to learn to differentiate between palpations of various pulse types, due to limited comparative learning time with established masters, and so normally it takes many years to master the art. The purpose of this study was to introduce an offline TCM skill evaluation and comparison system that makes available learning of palpation without the master's presence. We record patient's radial artery pulse using an existing pressure-based pulse acquisition system, then annotate it with teachers' evaluation when palpating the same patient, assigned as likelihood of it being each pulse type, e.g. wiry, slippery, hesitant. These training data were separated into per-doctor and per-skill databases for evaluation and comparison purposes, using the following novel procedure: each database was used as training data to a panel of time-series data-mining algorithms, driven by two validation tests, with the created training models evaluated in mean-squared-error. Each validation of the panel and training data yielded an array of error terms, and we chose one to quantitatively evaluate palpation techniques, giving way to compute self consistency and mutual-similarity across different practitioners and techniques. Our experiment of two practitioners and 396 per-processing samples yielded the following: one of the physicians has much higher value of self-consistency for all tested pulse types. Also, the two physicians have high similarity in how they palpate the slipper pulse (P) type, but very dissimilar for hesitant (H) type. This system of skill comparisons may be more broadly applied in places where supervised learning algorithms can detect and use meaningful features in the data; we chose a panel of algorithms previously shown to be effective for many time-series types, but specialized algorithms may be added to improve feature-specific aspect of evaluation.