以智慧型手錶的生理徵象監測建立急診醫護過勞示警

The application of smart watch monitoring to construct an overwork prediction and alarm model for emergency healthcare professionals

劉士宏(Shih-Hung Liu)

指導教授 ： 賴飛羆

國立臺灣大學/電機資訊學院/生醫電子與資訊學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202104525

摘要

因工作產生的疲勞一直是職場上值得關注的問題之一。此外，醫護人員的工作疲倦，不僅會影響醫護人員的健康，甚至有可能影響患者的安全。過去的研究方式，與過勞相關的研究多是在下班後，以問卷和量表的形式進行，而不是即時監測的方式。隨著科技的進步，穿戴式設備的發明提供了一種可行的解決方案，可以在不影響醫護人員日常工作的情況下，進行即時生理測量。同時，機器學習技術有了巨大的進步，並已應用於各個領域。在這樣的狀況下，我們才能嘗試建構一個可以即時監控過勞發生的警告系統。

此前瞻性觀察型研究於2021年3月10日至6月20日，在台大醫院急診室進行。納入研究的醫護人員會配給一支智慧型手錶 (ASUS VivoWatch SP)。這是一支消費級穿戴裝置，可以檢測心率和氧飽和度等多項生理測量值。此外參與者必須在每次工作前後，各完成一份多軸向疲勞量表。通過這種方式，我們可以找出有工作相關疲勞的醫護人員。接著我們利用量表及機器學習的方式，嘗試構建一個模型，用作即時工作相關疲勞的監控模式。

我們一共收集了1,542份有效的前後問卷。根據多軸向疲勞量表，有85人被判定有與工作相關的疲勞。在參與實驗的醫護人員中，有87.7%的人從事護理師的工作；以上班時間而言，47.7%的醫護人員於試驗期間上小夜班 (15:30~23:30)，44.5%的人員則是白班 (07:30~15:30)。我們使用了幾種目前最突出的模型 (State of the Arts) 的決策樹演算法進行建構。針對全體受試者，通過CatBoost分類器模型，在接收者操作特性曲線 (Receiver Operator Characteristic Curve, ROC) 的曲線下面積 (Area Under the Curve, AUC) 方面得到較好的表現0.838（95% CI：0.742 – 0.918）。而精確召回曲線下面積 (Area Under the Precision-Recall Curve, AUPRC)為0.527（95% CI：0.344 – 0.699）。除此之外，我們還對 35歲以下的護理師進行了次群組分析。在操作特性曲線下面積 (AUC) 得到更好的性能，其結果為0.928（95% CI：0.839 – 0.991），而精確召回曲線下面積 (AUPRC) 為0.781（95% CI：0.617 – 0.0.919）。在這個次群組分析，通過XGBoost得到比CatBoost分類器模型更好的結果，但此模組在回放到整體群組時，並不能得到更好的結果。
從穿戴式裝備萃取出的上百個特徵裡，我們利用了31個選定特徵，成功構建了一個機器學習模型。該模型能夠針對在急診室工作的醫護人員，進行與工作相關的疲勞風險進行分類。未來，我們可以將該工具應用在更多的急診人員上，有助於辨認出有工作相關疲勞風險的醫護人員，進而避免憾事發生。

關鍵字

過勞 ； 醫護人員 ； 穿戴式裝置 ； 機器學習 ； 即時監控

並列摘要

Work-related fatigue has always been one of the issues of concern in the workplace. Moreover, work-related fatigue among the medical personnel will have an impact on not only the health of the healthcare providers but also the patient safety. We found that the research related to overwork was mostly conducted in the form of questionnaires and scales that were completed after work, but not a real-time monitoring. With the advancement of technology, the invention of wearable devices provides a feasible solution to take the real-time physiologic measurements without an impact on the daily work of medical staff. Meanwhile, technique of machine learning has got an enormous improvement and been applied in all kinds of territory.

This prospective observational study was approved by the institutional review board (NTUH-REC No.: 202011024RIN) and conducted from March 10th to June 20th, 2021 at the Emergency Department in National Taiwan University Hospital. Medical personnel were provided with a smart watch (ASUS VivoWatch SP), a commercially available device which can detect several physiologic measurements like heart rate and oxygen saturation. Besides, participants had to complete a multidimensional fatigue inventory questionnaire before and after their daily work. By this we could find out the participants who had work-related fatigue. After that, we try to construct a machine learning model to be used as a real-time work-related fatigue monitoring.

A total of 1,542 effective before-and-after questionnaires were collected. Of them, 85 were thought to have work-related fatigue according to the multidimensional fatigue inventory questionnaire. Most of participants worked as a nurse (87.7%) and 47.7 percent of healthcare providers worked in the evening shift, 44.5 percent of them in the day shift. For overall population, the better performance in terms of Area Under the Curve (AUC) was 0.838 (95% CI: 0.742 – 0.918), which was achieved by the model of CatBoost classifier, while Area Under the Precision-Recall Curve (AUPRC) was 0.527 (95% CI: 0.344 – 0.699). We also performed a subgroup analysis focusing on the nurses who were under 35 of age. The better performance in terms of AUC was 0.928 (95% CI: 0.839 – 0.991) and AUPRC was 0.781 (95% CI: 0.617 – 0.0.919) which was achieved by the model of XGBoost classifier.

Using 31 selected features derived from a wearable device, we successfully built machine models which were able to classifying the risk of work-related fatigue in the circumstance of the emergency department. Implementation of this tool may be useful to identify the healthcare providers at risk of work-related fatigue.

並列關鍵字

work-related fatigue ； healthcare provider ； wearable device ； machine learning ； real-time monitoring

參考文獻

[1] G. M. H. Swaen, L. G. P. M. van Amelsvoort, U. Bültmann, and I. J. Kant, “Fatigue as a risk factor for being injured in an occupational accident: results from the Maastricht Cohort Study,” Occupational and Environmental Medicine, vol. 60, no. suppl 1, pp. i88–i92, Jun. 2003, doi: 10.1136/oem.60.suppl_1.i88.

[2] L. K. Barger et al., “Extended work shifts and the risk of motor vehicle crashes among interns,” N Engl J Med, vol. 352, no. 2, pp. 125–134, Jan. 2005, doi: 10.1056/NEJMoa041401.

[3] N. T. Ayas et al., “Extended Work Duration and the Risk of Self-reported Percutaneous Injuries in Interns,” JAMA, vol. 296, no. 9, pp. 1055–1062, Sep. 2006, doi: 10.1001/jama.296.9.1055.

[4] L. K. Barger et al., “Impact of Extended-Duration Shifts on Medical Errors, Adverse Events, and Attentional Failures,” PLOS Medicine, vol. 3, no. 12, p. e487, Dec. 2006, doi: 10.1371/journal.pmed.0030487.

[5] C. P. West, A. D. Tan, T. M. Habermann, J. A. Sloan, and T. D. Shanafelt, “Association of Resident Fatigue and Distress With Perceived Medical Errors,” JAMA, vol. 302, no. 12, pp. 1294–1300, Sep. 2009, doi: 10.1001/jama.2009.1389.