醫療機構服務品質評估系統研究：以類神經網路預測復健科健保申報之核減率為例

The Research of Medical Service Quality Evaluation: From the Viewpoint of Predicting Reclaimed Rate Using Neural Networks

徐弘正(Horng-Jeng Shyu) ； 邱昭彰(Chao-Chang Chiu) ； 徐培倫(Pei-Lun Hsu) ； 林美芳(Meei-Fang Lin)

《中華民國復健醫學會雜誌》 29卷4期 (2001/12) Pp. 185-193

https://doi.org/10.6315/JRMA.200112.0185

摘要

此研究之目的在於尋求申報資料與醫療院所行為之內含聯性，因復健申報資料不斷擴增，審查醫師的審查查負荷也逐漸加大；為了提昇專業審查品質及監控精確度，審查醫師必需採用客觀行為指標以利審查案件之篩選，而儘可能避免主觀人為判斷之誤差。若能借助監控系統做申報案件之事前審查，並初步歸類各醫療院所服務品質等級，較易達到監控的目的。一旦能掌握各醫療院所之服務品質後，即能輔導等級較差之醫院改善其服務品質，並且同時達到大幅節省醫療資源及保障全民醫療權益之目的。
本研究由復健申報資料中建立16個審查指標，並收集中醫健保局之復健科、神經內科、神經外科、骨科與整型外科等62家醫療院所申報資料中具復健治療之處方加以分析，並以85年7月至86年2月作為電腦類神經網路學習訓練及預測範圍。因健保局提供之核減率，是以整個醫院之各科綜合核減率，故將申報資料分單科診所、醫院與綜合醫院及診所等三組分析比較。所有的復健申報資料以核減率分三組(第一組為0%~3.5%、第二組為3.5%~7%、第三組為7%以上)，本研究以審查指標為電腦學習之輸入項目，以核減率分組為學習之產項目，組由類神經網路學習建立起核減率分組之預測模式，於診所預測正確率部分為92.31%，醫院組為74.07%，綜合組為70.00%，結果顯示診所之核減預測正確率高於醫院及綜合組之核減預測正確率，此原因可能為論所大都為一個醫師處方，而醫院為許多醫師處方可得之結果，故單一醫師之行為較易預測得知。另外本研究亦以統計方式預測核減率作為類神經網路預測之比較，其結果顯示以類神經網路預測之正確率均高於統計方式之核減預測正確率。
本研究由申報資料中建立了審查指標，並經由類神經網路系統預測模式證實審查指標之可用性，審查醫師可籍著這些審查指標，了解申報資料的整體狀態，如能有良好的資訊架構，審查醫師即可迅速的對申報異常型態有所反應。健保局己有良好的資料倉儲系統，如再結合審查指標及智慧型決策分析技巧，則可做出更有效率的審核方式，如此不僅能降低醫療浪費，更可增進醫療服務品。

關鍵字

行為指標 ； 服務品質 ； 類神經網路 ； 人工智慧

並列摘要

The goal of this study was to investigate the subtle interrelationships between the insurance claim’s data and clinic performance of medical institutes. Because the medical claims data increase drastically, the review efforts and processes become rather complex and inefficient. The physicians of review committee should have an enterprise view of the claims.
We create the performance indicators from consistent data & meta data established from the clinic performance of medical institutes. We adopted 16 outpatient review indicators. We collected rehabilitation claims in middle branch of NHIB since July 1996 till Feb 1997. The prediction model created using both the neural network and statistical methods. All the medical institutes are divided into three groups. Group A reclaimed rate ranges from 0 to 3.5%. Group B reclaimed rate is from 3.5% to 7%. Group C reclaimed rate is above 7%. We used the review indicators as learning inputs & the categories of reviewed claims rate as learning outputs. The neural network system was fed with these data & enforced to discern the mapping from inputs to the proposed output. The results of group prediction accuracy rate showed that the Neural Networks outperformed the statistical model. The group prediction accuracy rate was 92.31% for local clinic that consisting of only one physician, while 74.07% of accuracy rate for regional hospitals whose individual reclaimed rate were collected from several departments or physicians. The results showed that the prediction accuracy rate of clinics groups was better than that of the hospital groups.
This study applied neural network whose learning architecture was constructed using the genetic algorism to provide physicians the ability to access data and turn it into valuable information. This information can help physician of review committee to evaluate the performance of medical institutes. If we can produce a set of review indicators from the database of medical claims database, the combination of data warehouse and review indicators will deliver appropriate and cost-effective solutions in the review strategy.

並列關鍵字

review indicator ； service quality ； neural networks ； artificial intelligence