Authors: Jheniffer Raimundo¹; Thalita Ubiali¹; Betânia Eugenia Rodrigues da Silva¹; Rafael Ribeiro Santos¹; Edson Borin¹; Maria Francisca Colella-Santos¹
¹Universidade Estadual de Campinas (UNICAMP)
Background: The Frequency Following Response (FFR) provides crucial insights into subcortical auditory processing and fine temporal coding. While traditional statistics are standard for clinical analysis, Artificial Intelligence (AI) offers new possibilities for pattern recognition. This study aims to integrate traditional and computational FFR analyses to provide a comprehensive diagnostic framework for children with school difficulties or stroke sequelae.
Method: Evaluated 101 children (7–15 years old) divided into three groups: School Difficulties (GE-I, n=33), Stroke (GE-II, n=10), and Typical Development (GC, n=58). FFR was evoked using the /da/ syllable. Statistical analysis employed the Bootstrap technique (95% CI, p < 0.05). For the computational approach, a pipeline was developed in Python (PyTorch) using supervised and self-supervised machine learning (ML) models, with data partitioning by subject to prevent data leakage.
Results: Time-domain analysis revealed significantly higher amplitudes for peaks D (p=0.01) and E (p=0.02) in the stroke group (GE-II) compared to others. Frequency-domain analysis also showed higher magnitudes for GE-II. Conversely, machine learning models functioned primarily as “patient detectors” rather than “condition detectors,” identifying each subject’s unique biological signature. t-SNE visualizations demonstrated strong intra-class consistency (clustering by individual) but a lack of group-based clustering, suggesting that sound encoding is highly individualized and influenced by phonemic and linguistic processing.
Conclusion: While descriptive statistics effectively differentiated the groups by magnitude, ML revealed high signal stability and intra-class consistency. Despite sample size limitations for automated group categorization, AI exploration confirmed that FFR captures a reliable and persistent neural signature. This integrated approach highlights the potential of FFR as a tool for precision and longitudinal monitoring.
