Capsule DenseNet++: Enhanced autism detection framework with deep learning and reinforcement learning-based lifestyle recommendation.
Journal:
Computers in biology and medicine
PMID:
40120178
Abstract
Autism Spectrum Disorder (ASD) is a complex neurological condition that impairs the ability to interact, communicate, and behave. It is becoming increasingly prevalent worldwide, with an increase in the number of young children diagnosed with ASD in Saudi Arabia. Timely identification and customized interventions are essential for enhancing developmental outcomes. However, existing diagnostic approaches are subjective, limiting the cost-effectiveness of their utilization and the uniformity of their outcomes across different communities. In light of these concerns, this study presents a two-phase deep learning framework for autism detection with lifestyle advice using reinforcement learning. In the first phase, the proposed framework utilizes advanced multiscale statistical techniques for feature extraction, such as measures of central tendencies, variability indices, and percentiles, incorporated with the CosmoNest Optimizer, which is a hybrid of the African Vultures Optimization Algorithm and Butterfly Optimization Algorithm. For accurate ASD identification, these optimized features were classified using Capsule DenseNet++, an advanced deep learning model that increases feature representation efficiency and interpretability. In the second stage, we implement a personalized lifestyle recommendation system using the Proximal Policy Optimization (PPO) algorithm, a reinforcement learning algorithm. In the PPO approach, lifestyle decisions are sequential actions aimed at optimizing interventions, therapies, or daily activities for a given person. The PPO system dynamically learns and adapts recommendations over time to improve its effectiveness. The framework was developed in Python and tested on two datasets: autism screening data and ASD screening data for toddlers in Saudi Arabia. The performance of the detection model was recorded in terms of accuracy (99.2 % and 99.3 %, respectively), precision (98.5 % and 98.7 %, respectively), sensitivity (98.7 % and 98.9 %, respectively), and F1-score (99.1 % and 99.2 %, respectively), demonstrating its robustness for ASD detection across both datasets.