Artificial Intelligence Based Clustering Algorithm for Pulse Diagnosis

In traditional medicine, pulse palpation is a unique diagnostic technique focusing on identifying patterns of symptoms through subjective assessment of bio-signals. However, its reliability and objectivity have been questioned. We developed an artificial intelligence-based algorithm for clustering doctors' diagnostic results using unsupervised clustering techniques on pulse waveform signals.

Raw pulse signals were recorded from both wrists of healthy individuals and were then analyzed, with diagnoses provided by a Korean Medicine doctor. To measure pairwise pulse similarity, Dynamic Time Warping (DTW) was used, and Multidimensional Scaling (MDS) was applied for dimensionality reduction, enabling the clustering and validation of data-driven diagnostic patterns.

Our findings revealed discrepancies between traditional pulse diagnosis and automated diagnoses, yet the clustering algorithm showed high alignment between data-driven groupings and expert diagnoses. Notably, pulse signals from the left wrist had better alignment in several categories than those from the right wrist (cosine similarity: left hand 0.56 ± 0.13; right hand 0.54 ± 0.15). The "Floating-Sinking" pattern was particularly identifiable, achieving the highest Cosine similarity (0.83).

The results suggest significant alignment between data-driven pattern identification and expert diagnoses, especially for the "Floating-Sinking" pattern. Further refinement with diverse populations is necessary, but data-driven diagnostic tools hold potential for standardizing and quantifying traditional pulse diagnosis, moving it toward a scientifically robust practice.

Clinical Research Information Service KCT0007655 (registered on 2024-02-29).