Machine Learning Model Revolutionizes Early Autism Detection

AutMedAI, a new machine learning (ML) model, enhances early detection of Autism Spectrum Disorder (ASD) with minimal medical and background data. The recent breakthrough promises to enhance early diagnosis and intervention of ASD for better outcomes. The research conducted a retrospective analysis of the Simons Foundation Powering Autism Research for Knowledge (SPARK) database, version 8. The model was tested and validated on datasets from SPARK, version 10 and the Simons Simplex Collection (SSC) to ensure its accuracy and generalizability (1^✔ ✔Trusted Source
Machine Learning Prediction of Autism Spectrum Disorder From a Minimal Set of Medical and Background Information

Go to source).

How AutMedAI Was Developed

The study was conducted in Sweden with the approval from Swedish Ethical Committee by following the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) reporting guidelines.

All patients were consented and the data collection was approved by the ethical committee for the SPARK and SSC projects involving 30,660 individuals (15,330 with ASD and 15,330 without ASD) for analysis from 31 university-affiliated research clinics and online in 26 US states. 28 basic medical screening and background information before 24 months of age was collected and utilized for the development of the model.

The model was developed using four different ML algorithms—logistic regression, decision tree, random forest, and eXtreme Gradient Boosting (XGBoost). Performance Metrics- Accuracy, AUROC, sensitivity, specificity, positive predictive value (PPV), and F1 score were used.

AutMedAI’s Fabulous Results

Among the four algorithms, XGBoost model, later named as AutMedAI, showed high accuracy in diagnosing ASD and the model was tested on new datasets, giving good results across different age groups and both sexes.

High performance was observed for AutMedAI model with an average area under the receiver operating characteristics curve (AUROC) of 0.895 indicating high accuracy in differentiating children with ASD from the non-ASD children.

Advertisement

The model had 0.805 sensitivity, 0.829 specificity and 0.897 for PPV and was able to correctly diagnose ASD in 78.9% cases of total tested cases. Both new SPARK participants validation and SSC data yielded an AUROC of 0.790 and sensitivity of 0.680, proving the models efficiency on different data types.

The positive outcomes of AutMedAI prove the possibility to use its algorithm for the early diagnosis of ASD in children as a non-invasive technique. It signifies that ASD can be predicted with a small amount of information, allowing for earlier diagnosis, and early interventions which are so vital for the developmental outcomes.

Advertisement

Comparing the cases that were correctly and incorrectly classified using the model, the results showed that the correctly classified children with ASD were significantly more symptomatic with lower levels of cognition as compared to the children, who were incorrectly classified. This emphasises the versatility of the model. The efficacy of the model may differ for various populations and environments. It is necessary to conduct more validation and integrate with more diagnostic tools.

Reference:

1. Machine Learning Prediction of Autism Spectrum Disorder From a Minimal Set of Medical and Background Information - (https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2822394?)

Source-Medindia