Introduction
Autism Spectrum Disorder (ASD) diagnosis currently faces a significant bottleneck. Because there is no biological "lab test" for autism, clinicians must manually synthesize data from medical records, neurological exams, and behavioral assessments. This labor-intensive process often results in waiting lists ranging from 2 to 12 months. Given that early intervention is the most critical factor in improving long-term outcomes for children with ASD, there is an urgent need for automated screening tools that can process initial medical documentation and prioritize urgent cases.
System Architecture
The researchers developed a multi-stage machine learning pipeline designed to transform physical, handwritten medical records into actionable diagnostic signals. The process begins with digitization, where scanned forms undergo entropy-based de-skewing and Sobel operator-based edge detection for automated de-identification of sensitive patient information.
Following preprocessing, the system employs Handwriting Recognition to extract text contents, which are then passed to the ASD detection module. This module is divided into two primary phases: document representation and document classification.
Feature Extraction and Document Representation
A core contribution of this work is the comparison of different text representation techniques. While traditional methods like Bag-of-Words (BoW) and TF-IDF rely on word counts, they often struggle with the noisy and varied vocabulary found in medical notes.
To address this, the authors utilized Doc2vec, a deep learning-based approach that embeds entire documents into a continuous vector space. Unlike lexical features, Doc2vec captures the semantic context of clinical observations, allowing the model to understand the underlying meaning of a doctor's notes rather than just the specific words used.
Experimental Results
The study utilized a dataset of medical forms from 199 patients (56 ASD-positive, 143 non-ASD), including referral forms, parent/teacher questionnaires, and social worker notes. To handle the class imbalance, the researchers employed upsampling techniques to balance the training set.
The integration of Doc2vec embeddings and upsampling yielded the strongest results:
- Recall: 91.1%
- Accuracy: 83.4%
In a clinical screening context, the authors prioritized the F2 score, which weights recall more heavily than precision. This ensures that the system minimizes "False Negatives"—missing a child who truly has ASD—even at the cost of more "False Positives" that would be cleared during a formal evaluation.
Conclusion
By leveraging advanced NLP techniques like Doc2vec, this pipeline demonstrates that meaningful diagnostic signals can be extracted from messy, semi-structured medical data. While the current sample size is a limitation, the high recall rate suggests that such a system could effectively serve as a "first pass" screening tool. Future work aims to expand the binary classification into a severity index, allowing clinics to prioritize the most severe cases on their waiting lists and further accelerate the path to intervention.
Summary updated