運用人工智慧技術於資料探勘與分析以進行骨質疏鬆症疾病照護

Applying Artificial Intelligence Technologies for Data Mining and Analysis in Management of Osteoporosis

洪暐傑(Wei-Chieh Hung)

指導教授 ： 梁明在 ； 楊智惠

義守大學/理工學院/化學工程學系暨生物技術與化學工程研究所/博士(2021年)

摘要

台灣進入高齡社會，高比例老年人口同時有著高盛行率的骨質疏鬆症，而這類患者有著較高脆弱性骨折發生率，尤其髖骨、脊椎、腕骨等部位，一旦骨折發生，將增加死亡率與導致失能。為有良好之骨質疏鬆症照護品質，國際骨質疏鬆症基金會提出了以個案管理照護模式進行之骨折聯合照護服務，然而執行此照護模式需花費大量人力與物力，造成推行上之瓶頸。本研究以人工智慧之資訊科技導入醫療照護，以智慧化模式替代繁瑣人工作業，期盼能更順利推動個案管理照護模式，提升骨質疏鬆症病患之照護品質。

研究以人工智慧技術打造良好之個案管理系統，藉此提供臨床照護，內容分為三項子議題，前兩項議題在於人工智慧研發，做為個案管理系統智慧化之工具。第一項議題，由於脊椎的壓迫性骨折為常見骨質疏鬆症之影響部位，因此研發文字探勘模型，從放射科醫師的文字報告中主動辨識壓迫性骨折個案，為了減少人工標記訓練模型所需耗費的時間，本研究提出以向量和最小化文字資料抽樣方法，以此抽樣方法建立的文字探勘模型，辨識效果之接收者操作特徵曲線(Area under the receiver operating characteristics, AUROC)為0.952 (95%信賴區間：0.950,0.955；p < 0.001)，高於隨機抽樣方法的AUROC:0.900 (95%信賴區間：0.896,0.903；p < 0.001)。第二項議題，以機器學習建立骨質疏鬆症預測模型，辨識罹病高風險族群，並進行驗證，預測之骨密度T數值與黃金準則雙能量X光吸收儀檢查結果無顯著差異(平均絕對誤差:0.74;95%信賴區間：-0.09,0.02；p = 0.213)，另一模型則可直接良好預測是否為骨質疏鬆症(AUROC為0.755；95%信賴區間：0.729, 0.780；p <0.001)。第三項議題，以此個案管理平台於義大醫院進行骨折聯合照護服務後，與推動照護服務前比較，顯著提升照護成效，其中骨質疏鬆症藥物持續率自66.9%提高至71.6%，脆弱性骨折個案運動率從29.1%提升至54.7%，跌倒發生率自28.1%降低為16.2%，而出院後30日內再住院率自14.9%降低為6.3%。

結合人工智慧技術所建立之個案管理平台，可主動辨識病患，進而顯著改善臨床醫療對於骨質疏鬆症之照護成效，其中針對文字資料之抽樣技術進而快速建立文字探勘模型，以及建立較佳化預測模型與驗證等技術，未來更可運用在其他疾病照護，可從既有文字資料辨識已罹病之病患，而以預測模型找出罹病之高風險個案，進一步提供醫療服務。

關鍵字

none

並列摘要

Taiwan is an aged society including a high ratio of older populations with high prevalence of osteoporosis. Patients with osteoporosis are at a greater risk of suffering from fragility fractures especially in the hip, spine, and wrist, which may cause mortality or physical disability. For better care quality of patients with osteoporosis, the International Osteoporosis Foundation projected case management models to implement “Fracture Liaison Services.” However, applying these case management models demanded high manpower and material resources, which became the bottleneck of processing them. This study aimed to apply artificial intelligence (AI) technologies in clinical care and substitute trivially manual work for successfully implementing case management models to improve the care quality of patients with osteoporosis.

This study focused on creating a useful case management system with AI technologies as a tool in clinical care. This study included three main topics. First, vertebral compression fracture (VCF) was one of the most common presentations of osteoporosis. Thus, text mining models were created from radiology textual reports to identify patients with VCF. This study proposed a sampling method and minimized the vector sum to reduce the time consumed in manual automations for training models. The area under the receiver operating characteristics (AUROC) of a text mining model based on the sampling method of minimized vector sum was 0.952 (95% confidence intervals [CIs]: 0.950, 0.955; p < 0.001), which was better than that based on the random sampling method (AUROC: 0.900, 95% CIs: 0.896, 0.903; p < 0.001). Second, predictive models based on machine learning technologies were built and validated to identify high risk populations in osteoporosis. No significant difference was found between the predictive model and dual-energy X-ray absorptiometry exam, which is a gold standard exam in osteoporosis (mean absolute error: 0.74; 95% CIs: −0.09, 0.02; p = 0.213). Another predictive model can well classify patients with/without osteoporosis (AUROC: 0.729; 95% CIs: 0.729, 0.780; p < 0.001). Third, the outcomes of implementing fracture liaison services improved with this case management system. Persistent anti-osteoporotic medications increased from 66.9% to 71.6%. Fall rates decreased from 28.1% to 16.2% and exercise rates increased from 29.1% to 54.7% in patients with fragility fractures. The 30-day readmission rates after discharge reduced from 14.9% to 6.3%.

Actively identifying patients to improve the clinical management outcomes of osteoporosis was achieved by using this case management system with AI technologies. Sampling techniques in the free text data for fast established text mining models and building better predictive models with reliability validation could be applied in management of other diseases beyond osteoporosis. This case management system enabled identification of patients from textual data, such as medical records or medical examinations reports, and populations who are at a high risk as defined by predictive models. Further medical cares can be provided once the patients are identified.

並列關鍵字

none

參考文獻

1. National Statistics, R.O.C.(Taiwan). National Statistics, R.O.C.(Taiwan). https://www.stat.gov.tw/ Accessed May 22nd, 2021

2. Chen IJ, Chiang CY, Li YH, et al. Nationwide cohort study of hip fractures: time trends in the incidence rates and projections up to 2035. Osteoporos Int 26:681-688, 2015

3. Hawker G, Ridout R, Ricupero M, Jaglal S, Bogoch E. The impact of a simple fracture clinic intervention in improving the diagnosis and treatment of osteoporosis in fragility fracture patients. Osteoporos Int 14:171-178, 2003

4. Lorentzon M, Cummings SR. Osteoporosis: the evolution of a diagnosis. J Intern Med 277:650-661, 2015

5. Lu Y-C, Lin YC, Lin Y-K, Liu Y-J, Chang K-H, Chieng P-U, Chan WP. Prevalence of Osteoporosis and Low Bone Mass in Older Chinese Population Based on Bone Mineral Density at Multiple Skeletal Sites. Sci Rep 6:25206, 2016