MACHINE LEARNING–DRIVEN OPTIMIZATION OF ACUTE APPENDICITIS CARE IN CHILDREN: FROM DIAGNOSIS TO TREATMENT kacperstolarz2001@gmail.com

PW03-05

MACHINE LEARNING–DRIVEN OPTIMIZATION OF ACUTE APPENDICITIS CARE IN CHILDREN: FROM DIAGNOSIS TO TREATMENT

12

Sylwia Sanakiewicz sylwia.sanakiewicz@student.uj.edu.pl Jagiellonian University Medical College Kraków Poland

Urszula Zacharska urszula.pakulska@student.uj.edu.pl University of Oxford Oxford United Kingdom

Maria Gruba maria.gruba@gmail.com Jagiellonian University Medical College Department of Pediatric Surgery Kraków Poland

Wojciech Górecki wojciech.gorecki@uj.edu.pl Jagiellonian University Medical College Department of Pediatric Surgery Kraków Poland

Poland

Oral or Poster

Despite being one of the most common abdominal emergencies in children, there is still no specific biomarker for predicting acute appendicitis (AA). Diagnosis and choice of treatment remains essentially clinical, backed by laboratory tests and ultrasonography. However due to unusual clinical presentation, prompt decision about way of management can be difficult. Therefore machine learning (ML), which leverages large amounts of labeled data to extract statistical patterns predictive of a target variable is believed to enhance treatment planning.

Data of 627 patients with diagnosis of appendicitis and/or abdominal pain was collected. Independent variables included the patient's characteristics, signs, symptoms, blood morphology and ultrasonography findings. Data underwent preprocessing to ensure its suitability for analysis.The preprocessed data was divided into a training, testing and validation set. We utilized the XGBoost algorithm for continuous outcome prediction and Random Forest Classifiers for binary classification tasks. Trained models were evaluated using suitable metrics: AUC ROC and F1-Score. Analysis considered predictive models for binary response variables (appendicitis/no appendicitis, complicated/ uncomplicated, laparotomy/laparoscopy, surgical/conservative management) and continuous response variables (length of stay).

We created an online tool for research purposes only. Our model’s ability to tell apart cases of AA from a different abdominal pain was 94% and to distinguish simple from complex AA was at 91%. Main predictors for diagnosis were: surrounding tissue reaction, diameter of appendix and abdominal defense.

ML may provide a novel solution for diagnosing, planning treatment for children with appendicitis.