Sunday, Aug 3: 4:00 PM - 5:50 PM
4020
Contributed Papers
Music City Center
Room: CC-103B
Main Sponsor
Section on Statistical Learning and Data Science
Presentations
Periodontal pocket depth is a widely used biomarker for diagnosing risk of periodontal disease. However, pocket depth typically exhibits skewness and heavy-tailedness, and its relationship with clinical risk factors is often nonlinear. Motivated by periodontal studies, this paper develops a robust single-index modal regression framework for analyzing skewed and heavy-tailed data. Our method has the following novel features: (1) a flexible two-piece scale Student-t error distribution that generalizes both normal and two-piece scale normal distributions; (2) a deep neural network with guaranteed monotonicity constraints to estimate the unknown single-index function; and (3) theoretical guarantees, including model identifiability and a universal approximation theorem. Our single-index model combines the flexibility of neural networks and the two-piece scale Student-t distribution, delivering robust mode-based estimation that is resistant to outliers, while retaining clinical interpretability through parametric index coefficients. We demonstrate the performance of our method through simulation studies and an application to periodontal disease data from the HealthPartners Institute of Minnesota. The proposed methodology is implemented in the R package DNNSIM.
Keywords
deep neural network
single index model
skewness
heavy-tailed
modal regression
Many data-driven risk prediction models have been developed for analysing time-to-event data. However, choosing the most suitable model for accurate predictions in a specific medical application remains a challenge. Simulation enables effective comparison based on equal-sized datasets. This study provided a comprehensive evaluation of the survival prediction performance of random survival forests, eXtreme Gradient Boosting, deep neural networks (DeepSurv), and Cox proportional hazards (PH) model, using both simulated and real datasets. We assessed model performance using C-index and Integrated Brier Score. The evaluation was performed under varying sample sizes, censoring proportions, addition of noise variables, and in the presence of different types of model misspecification. All the models improved in predictive performance with larger sample sizes but declined with higher censoring and with increase in number of noise variables. Tree-based models demonstrated promising predictive performance compared to the Cox PH model and DeepSurv in the presence of misspecification and large number of noise variables. The Cox PH model performed well with larger sample sizes and fewer noise variables. It also performed well when the model was correctly specified or had only minor misspecification.
Keywords
Simulation
Machine Learning
Survival analysis
Prediction modelling
Stroke
Accurate decision support in healthcare often requires effective handling of uncertainty inherent in clinical data. Traditional fuzzy systems rely on manually crafted rule bases, which can be inflexible and labor-intensive to design. In this work, we propose an adaptive rule generation framework that integrates neural networks with fuzzy logic to automatically derive fuzzy rules from numerical health indicators and demographic variables. Our method utilizes learnable fuzzy membership functions to transform raw data into fuzzy representations and employs an attention mechanism to dynamically assign weights to the generated rules. The overall architecture consists of a fuzzy encoding layer, an adaptive rule generation module, and a subsequent prediction layer, enabling end-to-end training for regression tasks such as the prediction of Body Mass Index (BMI), weight, and waist circumference. By eliminating the reliance on manually defined rules, the proposed framework enhances model flexibility and interpretability, paving the way for more robust and data-driven decision support systems in healthcare.
Keywords
fuzzy system
healthcare prediction
attention mechanism
Chronic kidney disease (CKD) is a disease characterized by the gradual decline of kidney function over time due to which kidneys slowly lose their ability to filter waste and excess fluid from the blood. This causes a buildup of toxin in the body leading to kidney failure (End-Stage Renal Disease, ESRD) and other serious health complications. The prevalence of CKD has become a major concerning public health issue globally and in the United States, with its prevalence steadily increasing over the years. Early detection of CKD is crucial for effective management and treatment so that the progression of ESRD can be prevented or delayed thus reducing the overly expensive treatment cost. In this paper, we explore the use of machine learning (ML) techniques to predict Chronic Kidney Disease (CKD) based on the South Carolina Behavioral Risk Factor Surveillance System (BRFSS) dataset. However, the dataset is imbalanced, with a much smaller number of CKD cases compared to healthy individuals. The study compares different ML algorithms and tackles the challenges of imbalanced data.
Keywords
Machine Learning
Chronic kidney disease
imbalanced data
South Carolina
BRFSS
First Author
Dilli Bhatta, University of South Carolina Upstate
Presenting Author
Dilli Bhatta, University of South Carolina Upstate
Complex biological features like the microbiome and gene expressions mediate disease progression by influencing immune and metabolic processes. Understanding these mediation roles is crucial for disease pathogenesis and treatment. However, high-dimensional mediation analysis is challenging due to structural dependencies, correlations, and hierarchical relationships, such as microbial taxonomies and gene pathways. The many mediators also complicate conventional approaches.
We propose Med-CNN, a Convolutional Neural Network (CNN)-based model that integrates biological networks to estimate mediation effects. Network-specific CNN outputs are condensed into an Integrative Mediation Metric (IMM) to capture key biological information while handling high-dimensional data and non-linear interactions. Our model accommodates complex structures and improves interpretability in mediation analysis.
Through simulations, Med-CNN showed lower mediation effect biases compared to conventional methods. In a real data application, it identified a mediation effect between ethnicity and vaginal pH levels, demonstrating its robustness in analyzing high-dimensional mediators.
Keywords
High-dimensional
Mediation analysis
Microbiome
Deep learning
Deep learning (DL) has garnered increasing attention in time-to-event prediction due to its ability to model complex nonlinear relationships while offering greater flexibility than traditional methods. In this work, we propose a non-crossing quantile regression framework that estimates multiple quantiles of event time simultaneously in right-censored survival data while ensuring valid quantile ordering. Unlike existing approaches that rely on multilayer perceptrons (MLPs), we leverage Kolmogorov-Arnold Networks (KAN) for efficient function approximation and Transformers for capturing intricate feature dependencies through self attention. To provide theoretical insights, we establish upper bounds on the prediction error of our quantile estimators. We evaluate our framework on both simulated and real-world datasets, benchmarking its performance against existing quantile-based and non-quantile-based methods. As shown in these experiments, our DL framework can achieve 30% to 50% error reduction in real data analysis compared with various baseline methods.
Keywords
Deep Learning
Non-Crossing Quantile Regression
Time-to-Event Prediction
Transformer
KAN
Survival Analysis
Quantization-Aware Training (QAT) was integrated into a 1D 6-layer convolutional neural network (CNN) using the Sleep-EDF Database Expanded to detect sleep disorders by classifying sleep stages. This approach enables efficient on-device execution with limited memory while preserving performance despite quantization. QAT CNN consistently demonstrates reliable performance across all sleep stages, with its 8-bit and 16-bit quantized versions achieving high specificity (99.7%–100%) and lower sensitivity across most stages, and peak sensitivity at Stage R (100%) with lower specificity. These findings highlight QAT CNN's suitability for edge devices, ensuring stage-specific reliability, ie., sensitivity for Stage R and specificity for other stages. In contrast, non-QAT CNN exhibits lower sensitivity and higher specificity (82.1%–99%) across all stages and inconsistent performance in its quantized forms: 8-bit CNN shows high sensitivity and low specificity at Stage 3/4 and the reverse at other stages, while 16-bit CNN shows similar inconsistencies at Stage 1, instead. Overall, this QAT-guided CNN provides consistent and dependable performance for deploying quantized models on-device.
Keywords
Quantization Aware Training
CNN
on-device
Sleep-EDF Database Expanded
sleep disorder detection
QAT