MICHIGAN MEDICINE A Shiny App to Help Differentiate Tuberculosis from Crohn’s Disease Peter D.R. Higgins and Julajak Limsrivilai University of Michigan Mahidol University, Bangkok, Thailand @ibddoctor Slide deck: https://speakerdeck.com/higgi13425 

MICHIGAN MEDICINE The Problem of Small Intestine Ulceration 2 very important and likely diagnoses Healthy small intestine Crohn’s disease ulceration Intestinal TB ulceration Treat with immune suppression Treat with antimicrobials Wrong Rx – antimicrobials – very, very sick Wrong Rx – immune suppression – very, very sick or DEAD TREATMENT: WRONG DIAGNOSIS: 

MICHIGAN MEDICINE Changing Epidemiology • No Crohn’s disease in Asia until the 1960s • All small bowel ulceration assumed to be TB • Now Crohn’s disease is exploding in incidence in Asia, particularly in cities. Am J Gast 2008; 103; 3167-82. Lancet 2017; 390; 2769-78. 

MICHIGAN MEDICINE How to Treat? • Asia-Pacific GI Consensus Guideline on small bowel ulceration: • If unsure whether intestinal TB (ITB) vs. Crohn’s disease, Empirically treat with 4 anti-tuberculous therapies for 8-12 weeks. • If not improved, likely to be Crohn’s disease • During the 8-12 weeks, patients can experience worsening bleeding, weight loss, fatigue, intestinal blockages, nausea/vomiting, and intestinal perforation – may progress to surgery Ooi CJ , Hilmi I , Makharia GK et al. The Asia Pacific Consensus Statements on Crohn's Disease Part 1: definition, diagnosis and epidemiology. J Gastroenterol Hepatol 2015; 31: 45–55. 

MICHIGAN MEDICINE Can You Differentiate CD from TB? • Literature review • Lots of purported independent predictors (55) published • Sensitivities of single lab tests for diagnosis of Intestinal TB: • AFB stained : 5.3-37.5% • Caseous granuloma : 10-40% • Culture (6-8 weeks) : 23-45.8% • Tissue PCR : 36.4-67.9% • Performed meta-analysis for each of 55 predictors with metafor • 38 quality studies • 2,117 Crohn’s disease cases • 1,589 intestinal tuberculosis cases 

MICHIGAN MEDICINE Can You Differentiate CD from TB? 

MICHIGAN MEDICINE Can You Differentiate CD from TB? 

MICHIGAN MEDICINE Can You Differentiate CD from TB? 

MICHIGAN MEDICINE Convert Diagnostic ORs to Likelihood Ratios Convert Pretest Probability P0 to Post-test Probability P’ 

MICHIGAN MEDICINE Chaining multiple LRs Each LR for which we have data available The relative prevalence of ITB ( !"# !"#$%& ) Probability of ITB = ( ) ∗(,- . ∗,- / ∗⋯ ) .2( ) $(( ) ∗,- . ∗,- / ∗⋯ ) For findings without available results, the model defaults to an LR of 1 for that finding 

MICHIGAN MEDICINE External multicenter validation • 590 patients were included, 427 CD and 163 ITB • Thai cohort: 242 patients, 137 CD and 105 ITB • Hong Kong cohort: 348 patients, 290 CD and 58 ITB Bangkok Hong Kong 

MICHIGAN MEDICINE Model validation • The available data was applied to sequential models • Clinical model • Endoscopy model • Clinical + endoscopy model • Clinical + endoscopy + pathology model • The performance of the model was determined by • area under the curve of receiver operating characteristic curves • sensitivity, specificity, accuracy, false positive & negative rate 

MICHIGAN MEDICINE AuROC Clinical: 0.766 (0.723 – 0.809) Endoscopy: 0.786 (0.748 – 0.824) Clinical+Endo: 0.858 (0.825 – 0.892) Clinical+Endo+Patho: 0.903 (0.854 – 0.952) Model Performance 590 patients in Thailand and Hong Kong 427 Crohn’s Disease and 163 Intestinal TB 

MICHIGAN MEDICINE This Is Easier in a Shiny App! • https://www.pathology.med.umich.edu/shiny/tbcrohns/ - Robust to incomplete data - Computes estimates immediately - Allows user to evaluate the effects of additional (sometimes expensive) testing - Now frequently use in Asia 

MICHIGAN MEDICINE Limitations and Edge Cases • The model forces a dichotomous choice • Medicine is often not dichotomous • Patient with weight loss and hemoccult + stool • Strange-looking ulceration in small intestine on scope • Biopsies: ulceration, nonspecific inflammation, and healing • Model suggests more likely to be TB • All TB tests negative • ??? • Re-scope, more biopsies with depth – small bowel adenocarcinoma 

MICHIGAN MEDICINE Conclusions & Lessons Learned • Shiny apps can make models useful at the point of care • Can not be definitive • It helps to be robust to missing data • Always consider edge cases! • Manuscript: Limsrivilai, J., Am J Gast. 2017; 112: 415-427. https://www.ncbi.nlm.nih.gov/pubmed/28045023 • Shiny app: https://www.pathology.med.umich.edu/shiny/tbcrohns/ 

MICHIGAN MEDICINE Thank You @ibddoctor Slide deck: https://speakerdeck.com/higgi13425