Autism is a developmental disorder that is often identified by complications with social interaction and communication. Even though knowledge about it has increased and instruments such as the ADOS and ADI are available to assess the disorder, it is usually distinguished from observable symptoms, which makes it especially difficult to diagnose. A variety of concerns arise because of this issue, one of them being long-term symptoms, which can be mitigated with early screening and consequently early treatment. This research focuses on improving the diagnosis pipeline of autism by training and testing several state-of-the-art machine learning models with an Autism Spectrum Disorder dataset from the University of California, Irvine, and using machine learning methods to quantitatively identify the most significant indicators of autism in toddlers. We first designed a neural network classification model and Random Forests classification model, and then we trained/tested them to identify the presence of autism in toddlers. We also used feature selection through LightGBM parameter optimization to identify which physical characteristics are most significant in giving rise to autism. The outcome of this study was a highly accurate classifier for predicting the presence of autism and significant information on the importance of several physical characteristics in indicating autism. This enhanced diagnosis is critical as it leads to a more personalized and early-stage treatment, which can alleviate the effects of autism thereafter.