# Noninvasive prescreening of pediatric adenoid hypertrophy using quantified MFCC statistics and clinical features: development and external validation

## TL;DR

This study presents a noninvasive, low-cost method to screen for pediatric adenoid hypertrophy using voice features and clinical data, showing promising accuracy and potential for use in primary care.

## Contribution

A novel noninvasive screening model for adenoid hypertrophy combining MFCC voice statistics and clinical features, validated externally.

## Key findings

- The model achieved strong internal performance with AUC of 0.81 and AP of 0.84.
- External validation showed AUC of 0.79 and AP of 0.88, maintaining discrimination.
- SHAP analysis highlighted MFCC variability and airway-symptom variables as key contributors.

## Abstract

Adenoid hypertrophy (AH) is a leading cause of pediatric obstructive sleep apnea, yet first-line diagnostics are invasive or resource-intensive. We developed and externally tested a low-cost, noninvasive screening model that fuses quantitative voice features with routine clinical variables for pre-endoscopy and primary-care triage.

In a dual-center cross-sectional study (N = 202), Center 1 (Capital Center for Children’s Health, Capital Medical University, n = 161) served as the development cohort and Center 2 (College of Otolaryngology Head and Neck Surgery, The 6th Medical Center, National Clinical Research Center for Otolaryngologic Diseases, Chinese PLA General Hospital, Beijing, China, n = 41) as the independent external cohort. Children produced sustained/a/phonations; Mel-frequency cepstral coefficients (MFCCs) were summarized into fixed statistics and combined with readily available clinical information. Modeling used patient-level aggregation with stratified 10-fold cross-validation in development. The final classifier was selected by a joint criterion of AUC and average precision (AP), then a single Youden-derived locked cutoff was determined in the development set and applied unchanged to the external cohort. Discrimination (AUC/AP), calibration (Brier score, slope, intercept), and clinical utility were evaluated.

Internal performance was stable (AUC = 0.81; AP = 0.84). On the small external cohort, discrimination remained (AUC = 0.79; AP = 0.88). At the locked cutoff, the model achieved clinically actionable sensitivity/specificity with balanced F1. Calibration was acceptable (Brier = 0.20, slope = 0.71, intercept = 0.94). Decision-curve analysis showed positive net benefit across a wide range of threshold probabilities versus “treat-all” and “treat-none.” SHAP explainability indicated MFCC variability-related features and a subset of airway-symptom clinical variables as leading contributors, aligning with hyponasal resonance changes in AH.

A patient-level model with a locked decision threshold showed preservation of discrimination in a small external cohort, supporting a practical pathway for noninvasive, low-overhead AH triage prior to nasoendoscopy. Prospective multicenter studies are warranted.