Comprehensive Summary
This study aims to determine whether a hybrid LSTM (Long-Short Term Memory) and Attention model, along with a sliding window data preprocessing method would allow for more accurate diagnosis of autism. By using a lightweight hybrid model based on the LSTM and Attention mechanisms, along with a residual block channel combined with an Attention channel, feature fusion is allowed, and this method produces an accuracy of 81.1% on the HO brain atlas. Also, through the construction of a Top-N brain functional connectivity topological structure maps, patients and normal controls can be compared through calculation of their average Pearson correlation coefficient in varying brain atlases, improving the accuracy and reliability of autism identification from the dataset. The model proposed in this study was able to outperform single models that were based on either Attention or LSTM, with the best performance on AAL, CC200, CC400, DOS, and HO compared to other classification methods on ROI time series. Furthermore, through the Top-N brain functional connectivity topological structure maps, functional connections were identified, anatomical regions involved were evaluated, and their abnormalities aligned with established findings in ASD literature.
Outcomes and Implications
This research is relevant as autism diagnosis is currently on the rise in the US, a large concern as articles like ‘Is There An Autism Epidemic?’ published by John Hopkins Bloomberg School of Public Health proliferate in the media. It is well-known that early diagnosis of autism enhances the effectiveness of intervention strategies, and even more important is accurate diagnoses. Currently, the ADOS and ADI-R models are subjective and not generalizable, varying from physician to physician, but by the creation of an objective model with reproducible accuracy, autism diagnosis can become more reliable and allow for early intervention and better treatment outcomes.