Detecting autism from picture book narratives using deep neural utterance embeddings.
Aleksander Wawer, Izabela Chojnicka
PMID 35555933WHAT IT FOUND
Deep learning models classified autism from picture-book narratives better than two psychiatrists reading the same transcripts.
However, both models scored far below standard clinical tools like the ADOS-2. This is an exploratory proof-of-concept, not a diagnostic tool.
Key findings
01Computer-based models performed better than human raters at classifying participants based solely on transcribed narratives.
02Standardized instruments (ADOS-2 and SCQ) demonstrated diagnostic effectiveness way above the computer models.
03The best computer model (ELMo with page-level vectors) achieved sensitivity of 0.72 and specificity of 0.68.
STILL TO COME
How it was doneWhat they foundWhat it means for SLPs
Read the rest of this summary
You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.
What it does not show
The sample size was small (50 participants), limiting generalizability. The study only included individuals with ASD who had no intellectual disability (non-verbal IQ ≥ 80) and were fluent speakers, so results do not apply to those with more severe symptoms or language impairments. The models were tested on a specific narrative task (picture book description), so it is unclear if they would work for spontaneous conversation or other narrative types. The comparison to human raters was artificial; clinicians do not normally diagnose based solely on transcripts without observing the patient or gathering other information. The study was exploratory and did not validate the models on an independent dataset.
Declared interests
The authors declare no competing interests. The study was funded by the National Science Center of Poland and the Faculty of Psychology, University of Warsaw.
The easy way to misread this
Do not interpret the model's ability to outperform psychiatrists reading transcripts as evidence that AI can diagnose autism. The psychiatrists were restricted to text-only data, which is not how clinical diagnosis works. The AI models were still significantly less accurate than standard clinical tools like the ADOS-2.