Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterised by a wide range of clinical manifestations and developmental trajectories, which makes the diagnostic process complex and often challenging. Earlier identification of ASD enables earlier intervention and better outcomes. There is an increasing demand for reliable ASD screening tools that are easy and quick to administer. Recently, machine learning (ML) has been applied to improve the classification performance of first level screening tools, mainly parent-report questionnaires. In this study we used ML to improve the accuracy and applicability of a novel observational and interactive ASD screener, the Developmental Autism Early Screening (DAES). We previously developed this tool based on the Griffiths Scales of Child Development-third edition (Griffiths III) to intercept the early atypical developmental patterns in children with ASD risk, differentiating them from children with global developmental delay (DD) or neurotypical development (TD) in the first three years of life. In this study we explored and selected the potentially discriminative DAES items at two target age ranges (≤ 24 months and > 24 months) in a large sample of ASD, DD and TD children aged 12-48 months (n=610). We trained and tested five ML classifiers (random forest, RF, support vector classifier, SVC, decision tree, DT, logistic regression, LR, k-nearest neighbors, KNN) to classify ASD versus DD and TD children at the two different age ranges. RF and SVC were the two most effective algorithms for correctly detecting ASD children achieving very high accuracy (above 98%) from selections of 21 and 28 items (out of 36 DAES items) for children aged ≤ 24 and > 24 months, respectively. These findings confirm the validity of a shorter and faster version of the DAES using predictive items for specific age ranges. A widespread use of the tool will facilitate an earlier access to targeted intervention, allowing to redirect the atypical developmental trajectory towards a typical pathway in a greater number of children at risk of ASD.
Il Disturbo dello Spettro dell’Autismo (ASD) è un disturbo del neurosviluppo caratterizzato da un ampio range di manifestazioni cliniche e traiettorie evolutive, che rende complesso e spesso difficile il processo diagnostico. Una precoce identificazione dell’ASD può consentire l’inizio di un tempestivo intervento, con migliori outcomes. Vi è, pertanto, una crescente necessità di accurati strumenti di screening per l’autismo, rapidi e facili da somministrare. Recentemente, il machine learning (ML) è stato impiegato per migliorare le performances di strumenti di screening di primo livello, principalmente questionari compilati dai genitori. In questo studio, il ML è stato applicato per incrementare l’accuratezza e l’applicabilità di un nuovo strumento di screening osservazionale e interattivo per l’autismo, il DAES (Developmental Autism Early Screening). Abbiamo precedentemente sviluppato questo tool a partire dalla Griffith III (Griffiths Scales of Child Development – Third Edition), con l’obiettivo di intercettare precocemente patterns atipici di sviluppo nei bambini a rischio di ASD, distinguendoli da quelli con ritardo globale dello sviluppo (DD) o sviluppo tipico (TD), nei primi tre anni di vita. In questo studio abbiamo analizzato e selezionato gli items del DAES maggiormente discriminativi in due distinte fasce d’età (≤ 24 mesi e > 24 mesi), in un ampio campione di bambini (n = 610) con ASD, DD e TD, di età compresa tra 12 e 48 mesi. Sono stati addestrati e testati cinque algoritmi di ML (Random Forest – RF, Support Vector Classifier – SVC, Decision Tree – DT, Regressione Logistica – LR, e K-Nearest Neighbors – KNN) per classificare i bambini a rischio di ASD rispetto a quelli con DD e TD, nelle due differenti fasce d’età. RF e SVC sono risultati i modelli più efficaci, con un’accuratezza superiore al 98%, utilizzando i 21 e 28 items selezionati (a partire dai 36 items iniziali del DAES), rispettivamente nelle fasce d’età ≤ 24 mesi e > 24 mesi. Questi risultati confermano la validità di una versione più breve e rapida del DAES, basata su items predittivi specifici per fasce d’età. Un uso diffuso dello strumento potrà facilitare un più precoce accesso a interventi mirati, consentendo di reindirizzare positivamente la traiettoria atipica di sviluppo in un più ampio numero di bambini a rischio di ASD.
Design of a Machine Learning-based classifier for enhancing the accuracy and applicability of DAES, a novel autism screening tool [Sviluppo di un classificatore basato sul Machine Learning per migliorare l'accuratezza e l’applicabilità del DAES, un nuovo strumento di screening per l’autismo] / Cirnigliaro, L.. - (2026 Jan 28).
Design of a Machine Learning-based classifier for enhancing the accuracy and applicability of DAES, a novel autism screening tool [Sviluppo di un classificatore basato sul Machine Learning per migliorare l'accuratezza e l’applicabilità del DAES, un nuovo strumento di screening per l’autismo]
CIRNIGLIARO, LARA
2026-01-28
Abstract
Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterised by a wide range of clinical manifestations and developmental trajectories, which makes the diagnostic process complex and often challenging. Earlier identification of ASD enables earlier intervention and better outcomes. There is an increasing demand for reliable ASD screening tools that are easy and quick to administer. Recently, machine learning (ML) has been applied to improve the classification performance of first level screening tools, mainly parent-report questionnaires. In this study we used ML to improve the accuracy and applicability of a novel observational and interactive ASD screener, the Developmental Autism Early Screening (DAES). We previously developed this tool based on the Griffiths Scales of Child Development-third edition (Griffiths III) to intercept the early atypical developmental patterns in children with ASD risk, differentiating them from children with global developmental delay (DD) or neurotypical development (TD) in the first three years of life. In this study we explored and selected the potentially discriminative DAES items at two target age ranges (≤ 24 months and > 24 months) in a large sample of ASD, DD and TD children aged 12-48 months (n=610). We trained and tested five ML classifiers (random forest, RF, support vector classifier, SVC, decision tree, DT, logistic regression, LR, k-nearest neighbors, KNN) to classify ASD versus DD and TD children at the two different age ranges. RF and SVC were the two most effective algorithms for correctly detecting ASD children achieving very high accuracy (above 98%) from selections of 21 and 28 items (out of 36 DAES items) for children aged ≤ 24 and > 24 months, respectively. These findings confirm the validity of a shorter and faster version of the DAES using predictive items for specific age ranges. A widespread use of the tool will facilitate an earlier access to targeted intervention, allowing to redirect the atypical developmental trajectory towards a typical pathway in a greater number of children at risk of ASD.| File | Dimensione | Formato | |
|---|---|---|---|
|
TESI Dottorato Lara Cirnigliaro versione definitiva.pdf
accesso aperto
Tipologia:
Tesi di dottorato
Licenza:
PUBBLICO - Pubblico con Copyright
Dimensione
1.89 MB
Formato
Adobe PDF
|
1.89 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


