利用決策樹模型從免疫性血小班缺乏紫斑症病人中鑑別出較高機率進展成全身紅斑性狼瘡的族群：一個全國性世代研究

Identification of high probability of SLE development among ITP patients by means of decision-tree model: a nationwide cohort study

目的：為建立一個可有效且較低成本的決策樹模型，以鑑別出具高風險發展成全身紅斑性狼瘡的免疫性血小板缺乏紫斑症病患。方法：利用健保資料庫從1997年至2012年蒐集診斷為免疫性血小板缺乏紫斑症但未有全身紅斑性狼瘡的病人，並追蹤是否後續出現全身紅斑性狼瘡。同時分析病人的症狀、其他共病症以及平均口服類固醇用量，藉此建構出決策樹模型，以推估可能進展成全身紅斑性狼瘡的機率。結果：總計有12,265位免疫性血小板缺乏紫斑症病人納入此研究，而有80位病人後來出現全身紅斑性狼瘡。所有免疫性血小板缺乏紫斑症病人隨機分配至資料訓練組（共7,186位病人，包含57位全身紅斑性狼瘡患者）與測試驗證組（共3,079位病人，包括23位全身紅斑性狼瘡患者）；前者用來推導出決策樹模型，後者用來驗證前述模型，可以發現此決策樹模型具有高敏感性（78.2%）、特異性（99.2%）與陰性預測值（99.8%）。為使決策樹更易於應用，我們也提出另外不同複雜度的模型。結論：在僅知病人的症狀、其他共病症以及平均口服類固醇用量，而無其他實驗室數據的情況下，我們建構出不同的決策樹模型，藉以在不同的臨床狀況下鑑別出具有較高機率進展成全身紅斑性狼瘡的病人。

Objective: To determine a decision tree model that is an effective as well as less time- and cost-consuming method to identify those with high probability of development of systemic lupus erythematosus (SLE) amongst patients with immune thrombocytopenic purpura (ITP). Methods: We identified ITP patients without previous SLE diagnosis from the National Health Insurance Research Database between 1997 and 2012 and ascertained those who had the diagnosis of SLE during follow up. We also analyzed the symptoms and comorbidities as well as the dose of average oral steroid to derive the decision trees, which classified the ITP patients with different probabilities of development of SLE. Results: A total of 10,265 ITP patients were enrolled, among whom 80 patients developed SLE while following-up. The whole ITP patients were allocated to training group (7,186 patients including 57 with SLE) and testing group (3,079 patients including 23 with SLE); the former was used for derivation of the decision-tree based model and the latter for validation of the previously mentioned model, and provided high sensitivity (78.2%), specificity (99.2%) and negative prediction value (99.8%). To reduce the complexity, we also proposed another models with different complexity parameters (CP). Conclusion: We derived classification decision tree exempt from the necessity of laboratory data and suitable for various clinical scenarios of ITP patients, amongst whom were those with a high probability of SLE development.