背景
菌血症是造成病患死亡很重要的原因之一，雖然社區型菌血症的死亡率較低，但隨著高齡化社會的演進，社區的人口結構產生了變化，隨之而來的是，社區菌血症 的嚴重度產生很高的異質性，發生率也隨著人口結構老化而有上升的趨勢。若以多變 項羅吉斯回歸的統計方法，將需考量符合模型的假設以及變項之間潛在的重要交互作 用，而影響預測準確性。
目的
應用群集分析的方法於社區型菌血症的族群區別同質性的分群，再進一步利用決策樹方法分析在不同分群，影響 30 天死亡率的重要危險因子，以及適當抗生素使 用時間與 30 天死亡率的切點，並且比較與傳統羅吉斯回歸之預測準確度。
方法
本研究為單中心回溯性世代研究，2008.01.01-2013.12.31 的資料為 Modeling data
set，2019.01.01-2019.12.31 的資料為 Confirmation data set，皆納入急診採檢之血液培 養有陽性結果者，並且排除急診就醫前已有菌血症診斷者、護理之家的病人及沒有 30 天完整資料者。首先 Modeling data set 會進行群集分析，產生的分群結果會進一步個 別進行分類與回歸樹分析，並且與羅吉斯回歸進行比較，同時利用 Confirmation data set 進行外部效度的驗證，驗證 2019 年資料進行分群特徵配對後，使用對應的 CART 模型的 30 天死亡率預測準確度，本研究也會利用此分析方式了解不同分群中，重要 影響 30 天死亡率的適當抗生素使用時間點，以更全面了解分群間的差異特徵。
結果
Modeling data set 總共納入 2208 位病人進入群集分析，最理想的分群數為兩群，
Cluster 1(低死亡率組)有 1498 位病人，Cluster 2(高死亡率組)有 710 位病人。低 死亡率組的 30 天死亡率為 12.2%，高死亡率組的 30 天死亡率為 17.3%，兩者具有統 計學上的顯著差異(p<0.0012)。低死亡率組特徵為革蘭氏陰性菌(99%)，感染來源 為泌尿道感染(43%)、膽道感染(14%)及肝膿瘍(5%);高死亡率組特徵為革蘭氏 陽性菌(84%)，感染來源為皮膚軟組織感染(25%)、肺炎(19%)、骨關節(10%)、 心內膜炎(9%)及血管內感染(7%)。低死亡率組的分類與回歸樹節點依序為 PBS≧ 6 分à非泌尿道感染以及 PBS< 6 分à致死性共病症à敗血性休克à血糖<104mg/dL， 高死亡率組為 PBS≧6 分以及 PBS< 6 分à敗血性休克à惡性腫瘤。低死亡率組的決 策樹預測準確度 vs. 羅吉斯回歸為 0.8998 vs. 0.8886，高死亡率組的決策樹預測準確 度vs. 羅吉斯回歸為0.9061vs.0.8826，低死亡率組外部效度準確度為0.9061vs.0.895， 高死亡率組為 0.9109 vs. 0.9109。低死亡率族群如果有敗血性休克而無致死性共病症 時，則能否在進入急診 47 小時內使用上適當抗生素為重要的死亡影響因素，高死亡 率組如果 PBS<6 分且有敗血性休克，則能否在進入急診 13 小時內使用上適當抗生素為重要的死亡影響因素。
結論
社區型菌血症的病人主要有兩個特徵族群，低死亡率組主要以革蘭氏陰性菌為 主;高死亡率組主要以革蘭氏陽性菌為主。利用群集分析降低資料異質性，與傳統羅 吉斯回歸和單獨使用分類回歸樹相比，確實能有效提升模型的預測準確度。不同分群 理想的使用到適當抗生素時間為，低死亡率族群能否在進入急診 47 小時內使用到適 當抗生素為重要的死亡影響因素，高死亡率族群則是 13 小時內使用到適當抗生素為 重要的死亡影響因素。

Bacteremia was associated with high morbidity and among the leading cause of mortality. There was a fundamental structure change in the community in Taiwan due to aged society and home healthcare services. Current research on this topic mostly used multivariate logistic regression comparing the contribution of different variables. However, the heterogeneity of community-onset bacteremia led to relatively low positive predictive value among common clinical scoring systems and failed to precisely categorize different risk groups. Therefore, the aim of this study attempted to explore homogeneous groups of patients within community-onset bacteremia by using cluster analysis and select important risk factors for 30-day mortality in different clusters through classification and regression tree (CART). Thus, we conducted a retrospective cohort study. Modeling data set was from 2008 to 2013 and confirmation data set was 2019. Results of this study showed the optimal number of clusters was two clusters. Low mortality group was characterized by gram- negative bacteremia and the source of infection was urinary tract, biliary and liver abscess. There were five decision nodes for mortality 1. PBS≧6 2. Fatal comorbidity 3. Urinary tract infection 4. Septic shock 5. Glucose<104 mg/dL. The CART-derived time-to- appropriate antibiotic was 47 hours. High mortality group was characterized by gram- positive bacteremia and the source of infection was skin and soft-tissue, pneumonia, bone joint, infective endocarditis and vascular infection. The decision nodes were 1. PBS≧6 2. Septic shock 3. Malignancy. The CART-derived time-to-appropriate antibiotic was 13 hours. The accuracy compared between CART and logistic regression were 0.8998 vs. 0.8886 in low mortality group and 0.9061 vs. 0.8826 in high mortality group. To conclude, this study may be of importance in explaining the heterogeneity of community-onset bacteremia with two homogeneous groups and better understandings of different risk factors.

參考文獻
1. Bates, D.W., K.E. Pruess, and T.H. Lee, How Bad Are Bacteremia and Sepsis?: Outcomes in a Cohort With Suspected Bacteremia. Archives of Internal Medicine, 1995. 155(6): p. 593-598.
2. 國家發展委員會. 中華民國人口推估 (2020 至 2070 年 ). 2020; Available from: https://www.ndc.gov.tw/en/cp.aspx?n=2E5DCB04C64512CC.
3. Laupland, K.B., et al., Burden of community-onset bloodstream infection: a population-based assessment. Epidemiol Infect, 2007. 135(6): p. 1037-42.
4. Lee, C.C., et al., Age-Related Trends in Adults with Community-Onset Bacteremia. Antimicrob Agents Chemother, 2017. 61(12).
5. Henrard, S., N. Speybroeck, and C. Hermans, Classification and regression tree analysis vs. multivariable linear and logistic regression methods as statistical tools for studying haemophilia. Haemophilia, 2015. 21(6): p. 715-22.
6. Guilamet, M.C.V., et al., Cluster analysis to define distinct clinical phenotypes among septic patients with bloodstream infections. Medicine (Baltimore), 2019. 98(16): p. e15276.
7. Nelson, A.N., et al., Optimal duration of antimicrobial therapy for uncomplicated Gram-negative bloodstream infections. Infection, 2017. 45(5): p. 613-620.
8. Chotiprasitsakul, D., et al., Comparing the Outcomes of Adults With Enterobacteriaceae Bacteremia Receiving Short-Course Versus Prolonged-Course Antibiotic Therapy in a Multicenter, Propensity Score–Matched Cohort. Clinical Infectious Diseases, 2017. 66(2): p. 172-177.
9. Lee, C.C., et al., Clinical Benefit of Appropriate Empirical Fluoroquinolone Therapy for Adults with Community-Onset Bacteremia in Comparison with Third- Generation-Cephalosporin Therapy. Antimicrob Agents Chemother, 2017. 61(2).
10. Wang, M., et al., Thresher: determining the number of clusters while removing outliers. BMC Bioinformatics, 2018. 19(1): p. 9.
11. Seifert, H., The Clinical Importance of Microbiological Findings in the Diagnosis and Management of Bloodstream Infections. Clinical Infectious Diseases, 2009. 48(Supplement_4): p. S238-S245.
12. Laupland, K.B. and D.L. Church, Population-based epidemiology and microbiology of community-onset bloodstream infections. Clin Microbiol Rev, 2014. 27(4): p. 647- 64.
13. Takeshita, N., et al., Unique characteristics of community-onset healthcare-
69 associated bloodstream infections: a multi-centre prospective surveillance study of bloodstream infections in Japan. J Hosp Infect, 2017. 96(1): p. 29-34.
14. Diekema, D.J., et al., Epidemiology and outcome of nosocomial and community-onset bloodstream infection. J Clin Microbiol, 2003. 41(8): p. 3655-60.
15. Deen, J., et al., Community-acquired bacterial bloodstream infections in developing countries in south and southeast Asia: a systematic review. Lancet Infect Dis, 2012.12(6): p. 480-7.
16. Dantes, R.B., et al., Preventability of hospital onset bacteremia and fungemia: A pilot study of a potential healthcare-associated infection outcome measure. Infect Control Hosp Epidemiol, 2019. 40(3): p. 358-361.
17. Kao, C.H., et al., Isolated pathogens and clinical outcomes of adult bacteremia in the emergency department: a retrospective study in a tertiary Referral Center. J
Microbiol Immunol Infect, 2011. 44(3): p. 215-21.
18. Mahon, C.R., Lehman, D. C., & Manuselis, G. , Textbook of diagnostic
microbiology. 6th ed. 2019: Elsevier Saunders.
19. Hongsuwan, M., et al., Increasing incidence of hospital-acquired and healthcare-associated bacteremia in northeast Thailand: a multicenter surveillance study. PLoS One, 2014. 9(10): p. e109324.
20. Hilf, M., et al., Antibiotic therapy for Pseudomonas aeruginosa bacteremia: outcome correlations in a prospective study of 200 patients. Am J Med, 1989. 87(5): p. 540-6.
21. Al-Hasan, M.N., et al., Predictive scoring model of mortality in Gram-negative
bloodstream infection. Clin Microbiol Infect, 2013. 19(10): p. 948-54.
22. Battle, S.E., et al., Derivation of a quick Pitt bacteremia score to predict mortality in patients with Gram-negative bloodstream infection. Infection, 2019. 47(4): p. 571-578.
23. Chow, J.W. and V.L. Yu, Combination antibiotic therapy versus monotherapy for
gram-negative bacteraemia: a commentary. Int J Antimicrob Agents, 1999. 11(1): p.
7-12.
24. Hill, P.C., et al., Prospective study of 424 cases of Staphylococcus aureus
bacteraemia: determination of factors affecting incidence and mortality. Intern Med J, 2001. 31(2): p. 97-103.
25. Yu, V.L., et al., An international prospective study of pneumococcal bacteremia:
correlation with in vitro resistance, antibiotics administered, and clinical outcome.
Clin Infect Dis, 2003. 37(2): p. 230-7.
26. Vaquero-Herrero, M.P., et al., The Pitt Bacteremia Score, Charlson Comorbidity
Index and Chronic Disease Score are useful tools for the prediction of mortality in patients with Candida bloodstream infection. Mycoses, 2017. 60(10): p. 676-685.
27. Henderson, H., et al., The Pitt Bacteremia Score Predicts Mortality in Nonbacteremic Infections. Clinical infectious diseases : an official publication of the Infectious Diseases Society of America, 2020. 70(9): p. 1826-1833.
28. Lee, A., et al., Detection of bloodstream infections in adults: how many blood cultures are needed? J Clin Microbiol, 2007. 45(11): p. 3546-8.
29. Lee, C.C., et al., Beneficial effects of early empirical administration of appropriate antimicrobials on survival and defervescence in adults with community-onset bacteremia. Crit Care, 2019. 23(1): p. 363.
30. Liu, Y., et al., How to Read Articles That Use Machine Learning: Users’ Guides to the Medical Literature. JAMA, 2019. 322(18): p. 1806-1816.
31. Cios, K.J., et al., Unsupervised Learning: Clustering, in Data Mining: A Knowledge Discovery Approach, K.J. Cios, et al., Editors. 2007, Springer US: Boston, MA. p. 257-288.
32. Cios, K.J., et al., Unsupervised Learning: Association Rules, in Data Mining: A Knowledge Discovery Approach, K.J. Cios, et al., Editors. 2007, Springer US: Boston, MA. p. 289-306.
33. Huang, Z., Clustering large data sets with mixed numeric and categorical values. Proceedings of the First
Pacific Asia Knowledge Discovery and Data Mining Conference, Singapore: World Scientific. 1997a: p. pp. 21–34.
34. Huang, Z., Extensions to the k-Means Algorithm for Clustering Large Data Sets with Categorical Values. Data Min. Knowl. Discov., 1998. 2(3): p. 283–304.
35. Budiaji, W. and F. Leisch, Simple K-Medoids Partitioning Algorithm for Mixed Variable Data. Algorithms, 2019. 12(9): p. 177.
36. Kaufman, L.R., P.J., Finding Groups in Data. 1990, New York, NY, USA: John Wiley and Sons Inc.
37. Jenhani, I., N.B. Amor, and Z. Elouedi, Decision trees as possibilistic classifiers. International Journal of Approximate Reasoning, 2008. 48(3): p. 784-807.
38. Trivedi, S., Z.A. Pardos, and N.T. Heffernan, The utility of clustering in prediction tasks. arXiv preprint arXiv:1509.06163, 2015.
39. Kobayashi, D., et al., A predictive rule for mortality of inpatients with Staphylococcus aureus bacteraemia: A classification and regression tree analysis. Eur J Intern Med, 2014. 25(10): p. 914-8.
40. Quan, H., et al., Coding algorithms for defining comorbidities in ICD-9-CM and ICD-10 administrative data. Med Care, 2005. 43(11): p. 1130-9.
41. McCABE, W.R. and G.G. JACKSON, Gram-Negative Bacteremia: II. Clinical, Laboratory, and Therapeutic Observations. Archives of Internal Medicine, 1962.110(6): p. 856-864.
42. Chen, H., P. Cohen, and S. Chen, How Big is a Big Odds Ratio? Interpreting the
Magnitudes of Odds Ratios in Epidemiological Studies. Communications in Statistics- Simulation and Computation, 2010. 39(4): p. 860-864.
43. Moore, W.C., et al., Identification of asthma phenotypes using cluster analysis in the Severe Asthma Research Program. Am J Respir Crit Care Med, 2010. 181(4): p. 315-23.
44. Sweeney, T.E., et al., Unsupervised Analysis of Transcriptomics in Bacterial Sepsis Across Multiple Datasets Reveals Three Robust Clusters. Crit Care Med, 2018.46(6): p. 915-925.
45. Zhang, Z., et al., Deep learning-based clustering robustly identified two classes of sepsis with both prognostic and predictive values. EBioMedicine, 2020. 62.
46. Yang, C.Y., et al., Differential effects of inappropriate empirical antibiotic therapy in adults with community-onset gram-positive and gram-negative aerobe bacteremia. J Infect Chemother, 2020. 26(2): p. 222-229.
47. Corl, K.A., et al., Delay in Antibiotic Administration Is Associated With Mortality
Among Septic Shock Patients With Staphylococcus aureus Bacteremia. Crit Care
Med, 2020. 48(4): p. 525-532.
48. Falcone, M., et al., Time to appropriate antibiotic therapy is a predictor of outcome in patients with bloodstream infection caused by KPC-producing Klebsiella pneumoniae. Crit Care, 2020. 24(1): p. 29.
49. Bonine, N.G., et al., Impact of Delayed Appropriate Antibiotic Therapy on Patient Outcomes by Antibiotic Resistance Status From Serious Gram-negative Bacterial Infections. Am J Med Sci, 2019. 357(2): p. 103-110.
50. Lodise, T.P., et al., A systematic review of the association between delayed
appropriate therapy and mortality among patients hospitalized with infections due to Klebsiella pneumoniae or Escherichia coli: how long is too long? BMC Infect Dis, 2018. 18(1): p. 625.
51. Zasowski, E.J., et al., Time Is of the Essence: The Impact of Delayed Antibiotic Therapy on Patient Outcomes in Hospital-Onset Enterococcal Bloodstream Infections. Clin Infect Dis, 2016. 62(10): p. 1242-1250.
52. Lodise, T.P., Jr., et al., Predictors of 30-day mortality among patients with Pseudomonas aeruginosa bloodstream infections: impact of delayed appropriate antibiotic selection. Antimicrob Agents Chemother, 2007. 51(10): p. 3510-5.
53. Lodise, T.P., et al., Outcomes analysis of delayed antibiotic treatment for hospital- acquired Staphylococcus aureus bacteremia. Clin Infect Dis, 2003. 36(11): p. 1418- 23.
54. Peralta, G., et al., Altered blood glucose concentration is associated with risk of death among patients with community-acquired Gram-negative rod bacteremia. BMC Infect Dis, 2010. 10: p. 181.