José G. González Rodarte

Background: Acute appendicitis is the most common abdominal surgical emergency, with diagnostic uncertainty greatest in resource-limited settings. Objectives: To develop and internally validate an interpretable, laboratory-driven machine learning approach to assist clinical decision-making in suspected appendicitis, including diagnosis, perforation detection, and surgical severity stratification. Methods: A retrospective cohort of 246 patients with histopathologically confirmed appendicitis and 45 controls with similar abdominal pain was analyzed at a secondary-level hospital in Mexico. After cleaning and imputation, 41 laboratory variables were used to train three models: Random Forest for appendicitis detection and perforation identification, and Support Vector Machine for surgical severity stratification. Class imbalance was addressed with synthetic oversampling, and feature selection prioritized clinical interpretability. Results: Appendicitis detection achieved excellent discrimination (AUC = 0.94), correctly identifying 90% of cases, with 100% specificity. The perforation model reached 100% sensitivity (AUC = 0.875), prioritizing safe detection of high-risk cases, while severity stratification showed moderate performance (AUC = 0.721), correctly identifying 81% of complicated cases without imaging. Conclusions: Laboratory-based ML models accurately detected acute appendicitis and identified all perforated cases using routine data alone, while surgical severity stratification showed moderate discrimination in the absence of imaging. These findings demonstrate the feasibility of laboratory-driven decision support for early risk assessment in resource-limited emergency settings and support further external validation.

According to the World Health Organization (WHO) suicide is the third leading cause of death among individuals aged fifteen to twenty-nine years. This age group often coincides with higher education, where high levels of stress, anxiety and depression—factors that have a significant influence on suicide behavior—have been detected. The incidence is higher among medical students even during their professional practice. To address this issue, we developed machine learning models to perform a binary classification: $\mathbf{0}$= no suicidal behaviors, $\mathbf{1}$= suicidal behaviors. The dataset used in this study comprised 688 instances and 342 features, collected from an original self-reported questionnaire developed by experts and applied to medical students. The instrument includes items on socioeconomic status, substance use, trauma and related factors, and all responses were labeled by domain experts. The dataset was processed and transformed using techniques such as Frequency Encoding, Multi-Label Binarizer, Forward and Backward Selection. Two algorithms were trained—Support Vector Machines (SVM) and Random Forest (RF)—and evaluated, achieving an AUC (area under the curve) of 0.81 and 0.80, respectively.