A Pilot Study Using Machine-learning Algorithms and Wearable Technology for the Early Detection of Postoperative Complications After Cardiothoracic Surgery

Abstract

Objective: To evaluate whether a machine-learning algorithm (ie, the "NightSignal" algorithm) can be used for the detection of postoperative complications before symptom onset after cardiothoracic surgery.

Background: Methods that enable the early detection of postoperative complications after cardiothoracic surgery are needed.

Methods: This was a prospective observational cohort study conducted from July 2021 to February 2023 at a single academic tertiary care hospital. Patients aged 18 years or older scheduled to undergo cardiothoracic surgery were recruited. Study participants wore a Fitbit watch continuously for at least 1 week preoperatively and up to 90 days postoperatively. The ability of the NightSignal algorithm-which was previously developed for the early detection of Covid-19-to detect postoperative complications was evaluated. The primary outcomes were algorithm sensitivity and specificity for postoperative event detection.

Results: A total of 56 patients undergoing cardiothoracic surgery met the inclusion criteria, of which 24 (42.9%) underwent thoracic operations and 32 (57.1%) underwent cardiac operations. The median age was 62 (Interquartile range: 51-68) years and 30 (53.6%) patients were female. The NightSignal algorithm detected 17 of the 21 postoperative events at a median of 2 (Interquartile range: 1-3) days before symptom onset, representing a sensitivity of 81%. The specificity, negative predictive value, and positive predictive value of the algorithm for the detection of postoperative events were 75%, 97%, and 28%, respectively.

Conclusions: Machine-learning analysis of biometric data collected from wearable devices has the potential to detect postoperative complications-before symptom onset-after cardiothoracic surgery.

Figure 1.

Consort Diagram of Study Participants

Figure 2.

Time series of patients’ resting heart rates collected during the pre- and postoperative periods. Each panel represents a single patient’s resting heart rate. Figure 2A is an example of a true positive for a patient who underwent minimally invasive mitral valve repair and was discharged on POD 7. He reported shortness of breath for 3 days and presented to the emergency department with bilateral pulmonary emboli on POD 33. The date of symptom onset is indicated by the red circle (POD 30). The NightSignal algorithm first detected the postoperative clinical event on POD 29, as indicated by the red alert on POD 29. The gray-shaded region of the graph indicates postoperative days that were excluded from the analysis as they occurred more than 7 days after the date of the postoperative event. Figure 2B is an example of a true negative for a patient who underwent video-assisted thoracoscopic right middle lobectomy, was discharged on POD 2, and had no postoperative events. In both Figure 2A and 2B, POD 0 indicates the date of surgery. The green, yellow, and red dashed lines represent the different thresholds of the NightSignal algorithm. The solid black represents the patient’s average daily resting heart rate. Solid vertical yellow and red lines correspond to yellow and red alerts issued by the NightSignal algorithm. Red alerts during the immediate postoperative period were considered to reflect physiologic changes characteristic of normal recovery from the operation.

Figure 3.

Analysis of the performance of the NightSignal algorithm for postoperative event detection (A) and the time between the date the NightSignal algorithm first detected the postoperative event and the documented date of the postoperative event.