Machine learning and online tests identify autistic adults with 92% accuracy

A combination of an autism questionnaire and online tests of emotion recognition, perception, and mental skills distinguished autistic from non-autistic adults with 92% accuracy. These findings come from a new study published in the journal Translational Psychiatry.

Identifying autism in adulthood can be difficult, which is problematic as demand for adult assessments grows. Lengthy waiting lists often delay access to a proper diagnosis and professional support. Autism varies widely between people, and adults may have developed ways of managing or concealing characteristics that would otherwise be noticed. Screening questionnaires can help identify autistic traits, but their results depend on how people understand and report their own experiences.

Meanwhile, studies comparing autistic and non-autistic people on individual mental tasks have often produced small or inconsistent differences. The researchers wanted to examine whether a computer model could recognize a broader pattern across numerous measurements. Rather than looking for one defining difficulty, the model considered how performance across emotion recognition, attention, memory, mental flexibility, and sensory processing might fit together.

Led jointly by Erik Van der Burg and Robert M. Jertberg of Vrije Universiteit Amsterdam and the Amsterdam Public Health Research Institute, the team analyzed information from 552 Dutch adults. The group included 332 women and had an average age of about 38 years old. These individuals had participated in three previous online studies. The sample included 286 adults who reported a formal autism diagnosis and 266 without an autism diagnosis.

Participants completed online tasks assessing their ability to recognize emotions in faces and voices, combine information from sight and sound, and control automatic responses. They also completed tests that measured their ability to direct attention, remember spatial information, and switch flexibly between activities. The researchers extracted 54 measurements from these tasks. They then used a form of artificial intelligence called machine learning to classify the participants based on their scores.

Free daily newsletter

The autistic participants were older on average, and the groups contained different proportions of women and men. To account for this, the researchers repeated the analysis using 250 participants closely matched by age and gender. This provided a stricter test to see if task performance alone could predict a diagnosis.

Using the online tasks alone, the model reached 81.8% accuracy in the full sample. However, some of that performance reflected demographic differences like age. Accuracy was 74% in the age and gender matched sample. By comparison, a standard 28-item autism questionnaire alone reached a maximum accuracy of 83.2% in that same matched group.

Combining the questionnaire scores with the task performance produced the strongest result with 92% accuracy. The combined model correctly identified 88% of autistic participants and 96% of non-autistic participants. Reaction times during emotion recognition tasks were especially informative for the computer model. Measures of response control, short-term memory, and mental flexibility also contributed to the high accuracy.

Importantly, the model found useful patterns among measurements that had not shown obvious average differences when examined separately. As the authors noted, many individual test scores did not show statistically significant differences between the two groups. In other words, a single test result might reveal very little by itself. Yet, a person’s overall pattern across several tasks can still predict an autism diagnosis when the data is combined.

Some limitations to the research should be noted. For instance, the autistic group contained individuals with higher overall intelligence and lower support needs than the broader autistic population. The study also compared diagnosed autistic adults with people from the general public. It did not compare them to people referred to a clinic with other possible conditions. This means the computer model has not yet been tested on the type of patients a doctor might evaluate during an initial clinical assessment.

The study, “Finding the forest in the trees: Using machine learning and online cognitive and perceptual measures to predict adult autism diagnosis,” was authored by Erik Van der Burg, Robert M. Jertberg, Hilde M. Geurts, Bhismadev Chakrabarti, and Sander Begeer.