使用基本身體量測資料及機器學習方法偵測醫療資源缺乏地區之弱力症

Prediction of dynapenia by common medical devices in area lacking medical resource

張煒晟(Wei-Sheng Chang)

指導教授 ： 賴飛羆

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202201460

摘要

簡介：自2018年起，台灣被正式列為高齡化社會。肌少症是長者中最常見的肌肉疾病，其可能近一步地導致身體殘疾或死亡。儘管如此，診斷肌少症的醫療資源成本非常高且診斷過程耗時，在醫療資源缺乏之地區不易實施。本研究旨在減少傳統肌少症診斷的成本和時間。
方法：自 2019 年以來，我們招募了 1,406 名年齡在 20 歲以上的參與者。總共有 3,356 筆測量數據。我們利用了 3 種不同的機器學習技術和一個深度神經網絡來預測發生弱力症的風險。年齡、性別、血壓、BMI 和體重被我們作為輸入特徵。此外，我們添加了bitmask擴充特徵以幫助模型區分缺失值。在訓練和驗證過程中，使用嵌套交叉驗證來消除模型偏差並逼近真實世界的評估。最後，我們利用 Shap 來實施模型的解釋。
結果：經過 5 次嵌套交叉驗證後，有 5 個效能的結果。我們將5 個結果平均後作為最終的效能分數。具有bitmask特徵的 DNN 有最佳的分數，即 AUROC 0.8、F1 得分 0.84、PPV 0.81 和靈敏度 0.89。根據 Shap 的模型解釋，年齡、收縮壓和體重是對模型預測有很大影響的決定性特徵。
結論：我們開發了一種以低醫療成本預測肌少症風險的模型。模型在具有可接受的預測準確率下，可在缺乏醫療資源的地區作為快速篩測，以降低醫療成本。

關鍵字

弱力症 ； 肌少症 ； 機器學習 ； 深度學習

並列摘要

Introduction: Since 2018, Taiwan has officially been designated as an aged society. Sarcopenia is the most common muscle disease that occurs in elders which may lead to physical disability or death. Nonetheless, the medical cost of sarcopenia diagnosis is extravagant and the diagnosis process is time-consuming. We aimed to reduce the cost and time of traditional sarcopenia diagnosis.
Methods: We have recruited 1,406 participants with ages above 20 since 2019. There is a total of 3,356 measurement data across time and participants. We utilized 3 different machine learning skills and a deep neural network to predict the risk of dynapenia. Age, sex, blood pressure, BMI, and weight were used as our input features. In addition, we added bitmask augmented features to help models distinguish missing values. During training and validation, nested cross-validation was used to remove the model bias and approximate the real-world evaluation. Finally, we took advantage of Shap to practice the model interpretation.
Results: After 5-fold nested cross-validation, there were 5 performance outcomes. We then averaged 5 results as the final performance. DNN with bitmask features achieved the best performance which was AUROC 0.80, F1 score 0.84, PPV 0.81, and sensitivity 0.89. By Shap interpretation, age, systolic, and weight were determinant features that have a great influence on model prediction.
Conclusion: We developed a model predicting the risk of dynapenia with low medical cost. With acceptable performance, our model can be used as a rapid test in the area lacking medical resource to reduce cost.

並列關鍵字

dynapenia ； sarcopenia ； machine learning ； deep learning

參考文獻

[1] Chang KC, Hung SH, Hsueh SJ, et al. Development of the Houston–Apollo model for older people living in remote areas in Taiwan. Geriatrics Gerontology International. 2021;21: 299-305.

[2] Lin Y-H, Yen C-H, Kuo M-S, Hu M-H, Wang C-Y. The cutoff values of performance tests to separate community-dwelling older adults with and without physical disability. Taiwan Geriatr Gerontol. 2012;7: 160-174.

[3] Pongchaiyakul C, Limpawattana P, Kotruchin P, Rajatanavin R. Prevalence of sarcopenia and associated factors among Thai population. Journal of bone and mineral metabolism. 2013;31: 346-350.

[4] Mazocco L, Gonzalez MC, Barbosa-Silva TG, Chagas P. Sarcopenia in Brazilian rural and urban elderly women: Is there any difference? Nutrition. 2019;58: 120-124.

[5] Gao L, Jiang J, Yang M, Hao Q, Luo L, Dong B. Prevalence of sarcopenia and associated factors in Chinese community-dwelling elderly: comparison between rural and urban areas. Journal of the American Medical Directors Association. 2015;16: 1003. e1001-1003. e1006.