Master's Lecture in Applied Statistics - Jóhann Páll Hreinsson

Master's student: Jóhann Páll Hreinsson

Title: Comparison of logistic regression and gradient boosting methods for outcome after lower gastrointestinal bleeding

___________________________________________

Faculty: Faculty of Physical Sciences

Advisors: Einar Stefán Björnsson, Professor at the Faculty of Medicine and Sigrún Helga Lund, Associate Professor at the Centre of Public Health Sciences.

Examiner: Garðar Sveinbjörnsson, Statistician at deCODE Genetics.

Abstract

Background: Risk scores that identify which patients with lower gastrointestinal bleeding (LGIB) do not require hospital-based intervention are lacking.

Aims: To develop such a score with logistic regression (LR) and gradient boosting (GB), furthermore, to compare the two methods. 
Methods: A retrospective, population-based study including patients presenting to the emergency room (ER) with LGIB from 2010-2013. Hospital-based intervention was defined as blood transfusion, endoscopic hemostasis, arterial embolization or surgery. The cohort was split into train (70%) and test (30%) data. Train data were used to train the models and the test data used to validate them. 

Results: Overall, 581 patients presented 625 times to the ER, mean age 61 (±22), males 49%. Of train data patients, 72% did not require hospital-based intervention. Independent predictors of no hospital-based intervention in LR were systolic pressure ≥100mmHg (Odds ratio [OR] 4.9; 95% confidence interval [CI] 1.2-21), hemoglobin >12g/dL (OR 103; 95%CI 42-285), hemoglobin 10.5-12.0g/dL (OR 19; 7.4-53), no antiplatelets (OR 3.7; 95%CI 2.0-7.1), no anticoagulants (OR 2.2; 95%CI 0.96-5.1), pulse ≤100 (OR 2.9; 95%CI 1.3-6.7), and no visible bleeding in ER (OR 3.8; 95%CI 2,0-7,3). The most important predictors in GB were hemoglobin, systolic pressure, pulse and no visible bleeding in ER. Validation on the test data showed a negative predictive value and area-under-curve of 96% (95%ÖB 91-99%) and 0,83 for LR compared to 97% (92-99%) and 0,82 for GB, respectively. 

Conclusions: A new risk score has been developed for LGIB, the predictive ability of LR and GB were similar in this context.