Association of cerebrospinal fluid amphiphysin-1 levels with cognition and Alzheimer's disease pathology biomarkers.

Autism spectrum disorder is a behaviorally‐defined condition that can be reliably diagnosed by 24 months of age (Dawson et al., ; Kleinman et al., ; Lord et al., ; Pierce et al., ). Despite this, the median age of diagnosis remains approximately 4 years old, with variability across the country from around 3 to over 5‐1/2 years (Shaw et al., ). In some cases, autistic toddlers are enrolled in services for developmental delays or co‐occurring conditions prior to formally receiving their diagnosis. Indeed, a recent study found that implementing specific screening for autism within EI resulted in increased identification of autistic toddlers already receiving services, which was most pronounced for minoritized families and thereby reduced diagnostic disparities in the EI system (Sheldrick et al., ). Screening for autism is recommended by the US Preventive Services Task Force in order to improve access to EI (Siu et al., ), but conducting screening within the EI context would allow access to specialized services (both within EI and supplemental to it). Though there is some variability in the diagnostic process, best practices recommend utilization of a multidisciplinary (or interdisciplinary) team using standardized assessment tools with appropriate validity evidence, recognizing that the output of a standardized tool is not the same as a diagnosis—a diagnosis is based on a skilled diagnostician integrating all relevant data to determine case status (Bishop & Lord, ). Minimally, the diagnostic process should include a direct assessment, an interview with a knowledgeable informant, an assessment of functional strengths and weaknesses, and a determination regarding the amount of supports an individual would require. Integration of all these information sources is a necessary component of any diagnostic evaluation, yet is often insufficiently addressed in training and guidance. Machine learning techniques provide an opportunity to integrate multiple disparate but correlated features in a classification problem—which is exactly what diagnosticians are expected to be able to do. Classification algorithms can accelerate diagnostic efforts by identifying the most salient diagnostic features and reducing redundancy while maintaining high degrees of accuracy. Further, they can provide a predicted probability of a diagnosis that can support a diagnostician in their decision‐making process. Already machine learning algorithms are being used for diagnostic purposes, both to improve diagnostic measures (Wagner et al., ) and for data that are not part of routine diagnostic evaluations, such as eye tracking patterns or bodily movements (Jones et al., ; Simeoli et al., ). There are also examples of using item‐level data from autism diagnostic instruments to try to improve accuracy of the measures in the presence of co‐occurring conditions (e.g., Schulte‐Rüther et al., ). Aggregating data across measures that are routinely collected in the diagnostic process, however, remains an important area of growth.