Dados do Trabalho

Título

Machine learning helps to predict pediatric OSA based on clinical features

Introdução

Pediatric OSA affects 6 to 30% of children, with negative impacto n cardiovascular functions and neurocognitive development, beside quality of life. Diagnosis is based on polysomnography (PSG), screening questionnaires still with low accuracy. Machine learning (ML) algorithms might help to recognize moderate-severe OSA using clinical features, as well as tonsil size.

Objetivo

To evaluate sensitivity, specificity, and accuracy of ML data analysis of clinical features and tonsil size for the identification of moderate sever OSA.

Métodos

In this restrospective study, patient charts of children aged 2 to 12 years, submitted to a full night PSG were analysed, obtaining data as age, sex, OSA severity, snoring, apneas, mouth breathing, frequent awakenings, OSA18 questionnaire, and size of adenoids and tonsils (Brodsky scale). Children were divided in Group 1 No or mild OSA and Group 2 moderate severe OSA. Features were selected by Support Vector Machine (SVM) and a linear kernel, and Recursive Feature Elimination (REF). Data were divided in 12 models, 11 for trainning, 1 for test, the confusional matrix calculated sensitivity, specificity, accuracy , VPP and VPN for the correct identification. Fisrt round analysed clinical features, the second round included size of adenois and tonsils.

Resultados

111 children, mean age 5 years, were included, 55 composing Group 1 and 56 Group 2, both groups similar for age, gender. and tonsil size. The selected features were nasal secretion, choking, dificult deglution, anger, unattention, difficulty to wake, snoring, restless sleep, hyperactivity, enuresis, apneas. Clinical features showed sentitivity (0.667), specificity (0,833), accuracy (0.743). Including adenoid and tonsil size to clinical features reduced sensitivity (0.542), specificity (0.625) accuracy( 0.583).

Conclusões

ML improves prediction of moderate to severe OSA in children based on clinical features when compared to common screening questionnaires. Adenoid and Tonsil size are not good features to identify moderate severe OSA.

Palavras -chave

pediatric OSA, clinical features, machine learning Support vector machine