運用分類技術建構住院壓瘡模式之研究;Construction of a Hospitalized Pressure Ulcer Assesment Model Using Classification Technologies

[[abstract]]壓瘡發生率是臨床照護重要護理指標，國內外專家對於壓瘡預防不餘遺力，但住院壓瘡的發生率仍持續居高不下，主要在於壓瘡危險因子多且複雜度高，為避免住院病人因壓瘡而導致傷口疼痛、手術治療、延長住院天數、感染風險及嚴重甚至死亡，且增加醫療成本支出，準確掌握壓瘡高危險因子以事先著手預防壓瘡的發生。 國外壓瘡高危險評估量表包括有Braden量表、Norton量表、Gosnell量表及Waterlow量表，這些量表雖分別被研究採用，但其所建置的判定標準或界定切點，各自有其照護對象或照護體系適用度或可用性；因此本研究以資料探勘分類技術之決策樹、邏輯斯迴歸及隨機森林三種分類器建立住院病人適用之壓瘡預測模式，總共收集11838筆住院紀錄，實驗資料再將30組訓練樣本分別進行資料分析。隨後採用十摺交叉驗證法(10-fold cross-validation)驗證預測模式效能。 實驗結果顯示決策樹敏感性平均值為79.94%、邏輯斯敏感性平均值為75.81%及隨機森林敏感性平均值為84.48%。模型效能評估結果，以隨機森林所建構的預測模型具有較佳的分類效能。本研究中預測因素包括皮膚無完整、收縮壓高低、表達能力差、巴氏量表評分總分低及微血管填充時間大於2秒是壓瘡最具有影響力的危險因子。 本研究可提供護理人員作為住院病人於入院時預測是否壓瘡的參考，辨識壓瘡高危險因子，進而在臨床實務中提供預防措施，提升醫療照護及服務品質。 The incidence of pressure ulcer is one of the essential indicators of clinical care. Even many scholars and experts have aimed to prevent the adverse event of pressure ulcer, but the high incidence of hospitalized pressure ulcer was unfortunately commonly found in clinical practices. Therefore, identifying the risk factors and implement preventive interventions of pressure ulcer to avoid wound, pain, surgical treatment, prolonged hospitalization, infection, mortality and health expenditure increasing become more critical. We found that the Braden scale, Norton scale, Gosnell scale and Waterlow scale were broadly used for risk evaluation of pressure ulcer. However, these scales were established by the criteria of various patients and applicability or usability of the caring system. In this study, we use three classifiers of data exploration technology includes decision tree, logistic regression, and random forest to create the prediction model for hospitalized patients with a pressure ulcer. A total of 11838 medical records of hospitalization were collected and analyzed in 30 sets of training samples, and following with a 10-fold cross-validation was conducted to verify the performance of these prediction models. The results revealed that the sensitivity of the decision tree, logistic regression, and random forest were 79.94%, 75.81% and 84.48% respectively. It demonstrated that the random forest has better classification efficiency of constructing a predictive model. We also found that the predictive factors for impaired skin integrity, systolic pressure, poor expression, a low score in the Basel scale and microvascular filling time greater than 2 seconds are the most influential risk factors for pressure ulcers. This study provides the critical risk factors to caregivers for patient physical assessment to predict pressure ulcer incidence of hospitalized patients and to implement preventive actions in clinical practice. Overall, these findings enhance the improvement of the medical care quality and service