使用可穿戴設備資料、空氣品質資料和臨床評估資料 的大腸癌術後恢復機器學習預測模型

Abstract

背景：隨著醫學技術的進步，微創手術因為其傷口較小、住院時間較短，已漸漸取代傳統手術成為大腸癌的主要治療方式。而如何在此基礎上去進一步探討術後復原狀況成為了我們想研究的主題。 目的：本研究希望能透過臨床數據並結合病人的生活型態指標，並利用機器學習模型預測病人術後恢復情況。 方法：我們蒐集來自台大醫院及台大醫院新竹分院兩家醫院的85位大腸癌微創手術之病人資料。使用穿戴式裝置收集參加者術後的步數資料，根據其住所位置選擇最近的測站收集當日的空氣品質指標中的PM2.5當作環境資料，並結合其生理及臨床資料去做預測。我們使用術後有無併發症作為判斷術後恢復好壞之依據。利用四種資料處理方式、六種機器學習分類模型去預測參加者術後恢復情況，並以可解釋模型探討各個特徵。 結果：在各個資料集機器學習模型中，以第三個資料集的Random Forest和第四個資料集的XgBoost有較好的成果。以此結果近一步分析各個特徵，我們發現參加者每日步數超過兩千步及居住環境空氣中PM2.5較低都會使模型傾向預測為恢復好。 結論：結合病人的生理及臨床數據、步數資料及所在地之測站資料，經過資料前處理，並透過機器學習預測及可解釋性模型分析特徵對參加者術後恢復狀況之判斷。我們認為如果病人術後能每日有基礎運動量 (以步數為依據，每日須超過兩千步) 並處於PM2.5不高 (PM2.5副指標小於36) 之環境，有助於幫助術後恢復。

Background: With the advancement of medical technology, minimally invasive surgery has gradually replaced traditional surgery as the main treatment for colorectal cancer because of its smaller wound and shortened length of hospitalization. How to further investigate the postoperative recovery status on this basis is the subject of our research. Objective: The study hopes to use the physiological and clinical data combined with the patient's lifestyle indicators and use machine learning models to predict the patient's postoperative recovery. Method: Totally 85 patients, data had been collected from National Taiwan University Hospital and National Taiwan University Hospital HsinChu Branch, undergoing minimally invasive surgery for colorectal cancer. We use wearable devices to collect the participant's step count data, select the nearest station according to their location to collect the PM2.5 in the air quality index of the day as environmental data, and merge the two lifestyle indicators above with physiological and clinical data to predict. We use postoperative complications, which were defined by Clavien-Dindo classification, as the basis for judging the situation of postoperative recovery. Four data pre-processing methods, six machine learning classification models were used to predict, and each feature was discussed in an interpretable model SHAP. Result: In each data set and machine learning model, the performance of the third data set trained with Random Forest and the performance of the fourth data set trained with XgBoost are better. Based on the results, we found that the participants whose steps/day exceeded 2,000 and low sub-index of PM2.5 in their living environment, the model tended to predict a good recovery. Conclusion: Combined with the patient's physiological and clinical data, step counts and air quality data, through the data pre-processing, machine learning prediction and interpretable model analysis, the merged data are used to classify the participants' postoperative recovery status. We found that if the patient had a daily basic exercise (based on the number of steps, more than 2,000 steps per day) and is in an environment where PM2.5 is low (PM2.5 sub-index is less than 36), it will help Postoperative recovery.