Saturday, 14 March 2015

Boiling down ADOS for autism detection (again)

Today I want to direct your attention to the paper by Kosmicki and colleagues [1] (open-access) reporting that the use of "machine learning algorithms" could help "streamline ASD [autism spectrum disorder] risk detection and screening."

Regular readers of this blog might have already cottoned on to the fact that any talk about applying "computational and statistical methods" to autism screening and/or diagnosis can really mean only one person and research group: Dennis Wall from Stanford University. To quote from his institutional website on this area of research, the aim is to "evaluate the degree of redundancy of the ADOS and ADIR and if so determine whether a reduced set of uncorrelated features could correctly classify individuals with the same accuracy as the gold-standard diagnostic tests." ADOS and ADI-R by the way, are some, if not the, gold-standard schedules when it comes to the assessment of autism. The idea is that boiling down these respective schedules might save both time and resources when it comes to identifying those where a diagnosis of ASD is indicated. In case you'd like some history about this line of work, look no further than here...

The latest paper from the Wall group continues the research journey looking this time at modules 2 and 3 of the ADOS where previous work looked at module 1 (see here). In case you're not familiar with the concept of modules in ADOS, it's all about selecting the correct module according to verbal fluency (see here) where module 1 is for those who have very little or inconsistent phrase speech and modules 2 and 3 represent increasing phrase speech with also a little more focus on the use of age-appropriate props.

The results? Based on the development of 'classifiers' for each module, several machine learning algorithms were developed and tested (see here). One of the algorithms, ADTree, is by the way, the same classifier used in the previous module 1 ADOS work [2]. But ADTree did not perform best on this occasion: "The logistic regression classifier based on analysis of archival records from ADOS module 2 consisted of nine items, 67.86% fewer than the complete ADOS module 2, and performed with 98.81% sensitivity and 89.39% specificity in independent testing." Further: "The SVM module 3 classifier based on analysis of archived ADOS module 3 records consisted of 12 items, 57.14% fewer than the complete ADOS module 3, and performed with more than 97% sensitivity and specificity in testing."

The authors conclude: "These results support the notion that fewer behaviors when measured using machine learning tools can achieve high levels of accuracy in autism risk prediction."

Anyone who has either professional or personal experience of undertaking an ADOS will know that this is a highly specialised assessment schedule which often requires some time to complete. It's nothing like as time-consuming as the ADI but still, significant efforts and resources are needed to carry out the assessment and do so with skill and reliability (and maintain those all-important reliability stats). Wall et al have really started to shake the establishment when it comes to ADOS (and ADI) when asking just how much of the schedule is really needed to assess for autism/ASD. This on top of their other work talking about assessment 'triage' via YouTube videos using, horror of horrors, non-clinical raters (see here). I'm not saying that these approaches are ready for clinical practice; quite a bit more replicative work is required [3] including crossing geographical boundaries. But something like the idea that "mobile health approaches that ultimately enable individuals to receive more expedient care than is possible under the current paradigms" is a tantalising prospect.

So: Start by The Jam.


[1] Kosmicki JA. et al. Searching for a minimal set of behaviors for autism detection through feature selection-based machine learning. Transl Psychiatry. 2015 Feb 24;5:e514.

[2] Wall DP. et al. Use of machine learning to shorten observation-based screening and diagnosis of autism. Transl Psychiatry. 2012 Apr 10;2:e100.

[3] Bone D. et al. Applying Machine Learning to Facilitate Autism Diagnostics: Pitfalls and Promises. J Autism Dev Disord. 2014 Oct 8.

---------- Kosmicki JA, Sochat V, Duda M, & Wall DP (2015). Searching for a minimal set of behaviors for autism detection through feature selection-based machine learning. Translational psychiatry, 5 PMID: 25710120