FRISTS: A Novel AI Framework for Interpretable Heart Failure Prediction Through AI Feature Selection and Time Series Recurrent Neural Networks

Overview

Heart disease is the leading cause of death in the United States since 1921. As heart failure (HF) is incurable and lethal, developing prediction models is crucial for prevention and early intervention. Current research struggles with two key challenges: (1) lackluster accuracy and precision when tested on real-world health records and (2) uninterpretable characteristics that confuse healthcare practitioners and prevent widespread adoption. To address these challenges, I introduce FRISTS (Feature Ranked Int erpretable Sequential Time Series), a novel prediction approach that achieves both high performance on real -world data and transparency in its decision -making. FRISTS leverages a new combination of sequential times series-based recurrent neural networks (R NN), e.g. long short -term memory (LSTM) networks, and AI - guided feature selection. Rigorous real-world testing demonstrates that FRISTS surpasses state-of-the-art baselines, outperforming the highest -accuracy (XGBoost, LSTMs) machine learning techniques in the literature. When evaluated on nearly 18 million electronic health records (EHR) in the Cerner Health Facts database, FRISTS yields an average F1 score of 0.805 and receiver operating characteristic area under the curve (ROC AUC) value of 0.990, a four-fold increase in performance compared to baselines. A SHAP- inspired (Shapley Additive Explanations) permutation method enables interpretable feature ranking, giving healthcare practitioners insight into the model’s decision-making and demonstrating that FRISTS captures more HF -related features than other interpretable models (Random Forest, Logistic Regression, and Decision Trees). Since FRISTS is extendable to any prediction task on health records, it accelerates the adoption of machine learning methods that achieve both real -world accuracy and interpretability for AI - assisted healthcare and disease prevention. SpecuSafe: A Non-invasive and Intelligent Approach to Cervical Examinations Aashritha Penumudi Thomas Jefferson High School for Science and Technology, Alexandria, VA Cervical cancer, a leading cause of death in low- and middle-class areas, can be identified using visual inspection with acetic acid, but such speculum-based inspections can be subjective in areas without medical proficiency and a triggering process for women with a history of sexual abuse. The goal was to develop a diagnosis method to aid health professionals in low-access areas by designing an advanced prototype speculum and website with an integrated high-accuracy image-classification model. This computer-based project utilized a requested dataset of cervix images from the World Health Organization International Agency for Research on Cancer (WHO IARC) labeled with cancer status and characteristics and was used to train a supervised image classification model. The images were processed by image segmentation and masks to reduce the influence of specular reflection. The data was then augmented to diversify the data and improve prediction abilities, then run through the model which is integrated into an instructive website. SpecuSafe is a novel model speculum with integrated technologies to ease self- speculum-insertion, pH testing, lighting, and imaging. It is rechargeable with a resin-finish and smaller length/radius for comfort. The machine learning model functions with a high validation accuracy of 83%. The SpecuSafe speculum and website can save lives by detecting cervical cancer early in vulnerable populations, potentially improving with future steps of expanding the dataset and directly feeding an endoscopic camera’s output to the model.