JMIR Publications

ABSTRACT

Background:

We attempted to quantitatively express changes in perioperative patient behavior using machine learning and triaxial acceleration data, and establish an index that can express differences in surgical invasiveness.

Objective:

Two patients who underwent invasive thoracoscopic surgery were included.

Methods:

The acceleration data were collected using a sensor placed on the chest. Four actions (walking, standing, sitting, and lying down) were identified using supervised machine learning, and the time and appearance probabilities for each action were calculated. Four two-dimensional vectors having with probabilities as coordinates on the axis were defined. The center of the four vectors was then calculated as a new index of behavioral patterns (iBP). Daily iBP was plotted, and the area surrounded by the points and the distance between the points were calculated to evaluate the changes in the indices.

Results:

Subject 1 underwent lung lobectomy and subject 2 underwent tumor biopsy. The lying time in subject 1 was prolonged immediately after surgery, followed by a decrease (preoperation, 11.5 [min/h]; postoperative day 1, 27.5; day 3, 13.1). In contrast, the lying time of subject 2 remained unchanged (17.9, 12.7, and 18.7 ). The enclosed areas were 0.0765 and 0.0036, and the distances were 1.13 and 0.47 in Subject 1 and Subject 2, respectively.

Conclusions:

We numerically expressed the behavioral patterns of surgical patients using the estimated action time. The new index facilitates the understanding of changes in perioperative behavioral patterns in a subject on a graph and the results reflect the differences in surgical invasiveness among the subjects. Clinical Trial: This study was approved by our Institutional Ethics Committee and registered with the University Hospital Medical Information Network Individual Case Data Repository (UMIN000026843).