Empathy Is a Stress Response - Choose Compassion instead
Influenza-like illness by National Health Insurance databases
1. Comorbidity Attributes that Link to
Unfavorable Outcomes of Influenza-
like Illness-related Inpatients -- A
Nationwide Cohort Analysis
以台灣全民健保資料庫分析探索類
流感關聯病人不良預後之共病特質
及預測模式
陳勁辰 金傳春 …
2. Introduction
• Susceptible –(1)-> Influenza –(2)-> Outcome
• Comorbidity (Como) affects (2)
• Real world: symptoms/syndrome groups
• Influenza-like illness (ILI)
• Aim: Como effect of ILI on outcomes
3. Methods
• National Health Insurance Database (NHID)
• Materials: One-million samples of NHID of
2007, 2008, 2009, 2010
• Roughly: Seasonal in 2007+2008; Pandemic in
2009+2010
• ILI defined by EID
• ILI-related inpatients: hospitalized with ILI or
ambulatory visits for ILI =< 1 day
4. Marsden-Haug N, Foster VB, Gould PL, Elbert E, Wang H, Pavlin JA. Code-based syndromic
surveillance for influenzalike illness by International Classification of Diseases, Ninth
Revision. Emerging Infect. Dis. 2007;13(2):207–216.
5. Methods
• ILIR cohort: EID definition
• Como: Selected by relevance and advisors
• Outcomes:
– Cost: from NHID
– Length of stay (LOS): from NHID
– Daily cost: Cost/LOS
– Death: by endpoint code
– ICU: by treatment code
– Adverse event (AE): Death or ICU
7. Strata by Age
• By social-economic status
• [0, 6): pre-school
• [6, 15): school children
• [15, 25): the youth
• [25, 45): young adults
• [45, 65): middle-aged
• [65, oo): the elderly
8. Methods
• Data management by SAS 9.3, SAS-SQL
programs
• “Big Data” computation on NTU virtual
machine remotely
• Statistics by StataMP 13.1 Mac
• Programs and files shared in “Clouds”
9. Models
• Outcome Y
• Como X
• Adjusted by Sex
• Y = B0 + B1 X + B2 Sex (Each Age stratum)
• Cost, LOS, Dcost -> log-transformed -> Linear
regression
• Death, ICU, AE -> Logistic regression
10. Algorithms
• ILI hetero each year? Checked by sex, age,
como’s, outcomes, ILI top-ten codes;
• Combine homo years, then proceed;
• In each age stratum, como is selected if:
– Sig in all regression models (all endpts)
– Age-specific prevalence>5%
11. Algorithm
• ILI Score = Sum of <Como> * <In Age>
• Internal validation by modeling ILI score on
outcomes
– Cost/LOS/Dcost by Spearman correlation
– Death/ICU/AE by ROC
13. Results: Yearly Comparison
• ILI ICD9: freq rank; fisher's exact test p=0.9090
• Sex: fisher's exact test p=0.2380
• Age: Mann-Whitney U test as scale; Fisher's exact
test as strata nominal; p<0.001 (age up with year)
• Como: each p<0.001 except preg, cong, imdef,
autoimm
• Endpt: each p<0.001 (cost, los, daycost, die, icu,
ae)
14. Results: Data Merge by Pandemic State
• Yearly data cannot be merged into one;
• Yearly formulae are not practical;
• Formulae by pandemic attribute (pan=0 or 1)
• Seasonal (pan=0): 2007+2008
• Pandemic (pan=1): 2009+2010