Identifying the Most Salient Audio and Language Features for Pediatric Specific Language Impairment Classification

Abstract

Specific language impairment, also known as SLI, is a pediatric language disorder that delays the development of typical speech functions without the influence of other developmental delays or neurological disorders. SLI prevents children from clearly communicating their thoughts or desires with others and can persist throughout their lives if left undiagnosed. With the ability to provide scalable diagnostic services in the comfort of one’s home, machine learning solutions offer the potential for an accessible screening method for SLI, enabling a parent or guardian to identify potential markers and consult with a speech and language therapist about clinical actions. To address this opportunity, I developed a machine-learning solution to classify SLI based on audio and language features derived from the Talkbank Collection of the CHILDES dataset. I applied feature selection to identify the most salient features using top-ranked gradient-boosting features, logistic regression coefficients, and mutual information scores. The gradient-boosting classifier outperformed the other two methods, achieving 85% average accuracy, 85% average precision, and 83% average recall. The top features across the three feature selection strategies were the z-score of mean utterance length, age, perplexity of 1-gram SLI, word types to word token ratio, number of nouns followed immediately by a verb, flesch–kincaid score, repetitions, possessives, and the z-score of word errors. Of note, the flesch-kincaid score and perplexity of n-gram sequences, while not new, are relatively understudied features in SLI analysis and would benefit from additional research. Interestingly, prior ML studies have found these features appear in the context of other conditions, such as mild cognitive impairment and dementia.