Development of Deep Ensembles to Screen for Autism and Symptom Severity Using Retinal Photographs

Automatic Cry Analysis: Deep Learning for Screening of Autism Spectrum Disorder in Early Childhood

PURPOSE

The objective of this study is to identify the acoustic characteristics of cries of Typically Developing (TD) and Autism Spectrum Disorder (ASD) children via Deep Learning (DL) techniques to support clinicians in the early detection of ASD.

METHODS

We used an existing cry dataset that included 31 children with ASD and 31 TD children aged between 18 and 54 months. Statistical analysis was applied to find differences between groups for different voice acoustic features such as jitter, shimmer and harmonics-to-noise ratio (HNR). A DL model based on Recursive Convolutional Neural Networks (R-CNN) was developed to classify cries of ASD and TD children.

RESULTS

We found a statistical significant increase in jitter and shimmer for ASD cries compared to TD, as well as a decrease in HNR for ASD cries. Additionally, the DL algorithm achieved an accuracy of 90.28% in differentiating ASD cries from TD.

CONCLUSION

Empowering clinicians with automatic non-invasive Artificial Intelligence (AI) tools based on cry vocal biomarkers holds considerable promise in advancing early detection and intervention initiatives for children at risk of ASD, thereby improving their developmental trajectories.

Oculomics approaches using retinal imaging to predict mental health disorders: a systematic review and meta-analysis

OBJECTIVE

This meta-analysis evaluated the diagnostic performance of oculomics approaches, including deep learning, machine learning, and logistic regression models, in detecting major mental disorders using retinal imaging.

METHODS

A systematic review identified 11 studies for inclusion. Study quality was assessed using the QUADAS-2 tool, revealing a high risk of bias, particularly in patient selection and index test design. Pooled sensitivity and specificity were estimated using random-effects models, and diagnostic performance was evaluated through a summary receiver operating characteristic curve.

RESULTS

The analysis included 13 diagnostic models across 11 studies, covering major depressive disorder, bipolar disorder, schizophrenia, obsessive-compulsive disorder, and autism spectrum disorder using color fundus photography, and optical coherence tomography (OCT), and OCT angiography. The pooled sensitivity was 0.89 (95% CI: 0.78-0.94), and specificity was 0.87 (95% CI: 0.74-0.95). The pooled area under the curve was 0.904, indicating high diagnostic accuracy. However, all studies exhibited a high risk of bias, primarily due to case-control study designs, lack of external validation, and selection bias in 77% of studies. Some models showed signs of overfitting, likely due to small sample sizes, insufficient validation, or dataset limitations. Additionally, no distinct retinal patterns specific to mental disorders were identified.

CONCLUSION

While oculomics demonstrates potential for detecting mental disorders through retinal imaging, significant methodological limitations, including high bias, overfitting risks, and the absence of disease-specific retinal biomarkers, limit its current clinical applicability. Future research should focus on large-scale, externally validated studies with prospective designs to establish reliable retinal markers for psychiatric diagnosis.

Retinal fundus imaging as biomarker for ADHD using machine learning for screening and visual attention stratification

Attention-deficit/hyperactivity disorder (ADHD), characterized by diagnostic complexity and symptom heterogeneity, is a prevalent neurodevelopmental disorder. Here, we explored the machine learning (ML) analysis of retinal fundus photographs as a noninvasive biomarker for ADHD screening and stratification of executive function (EF) deficits. From April to October 2022, 323 children and adolescents with ADHD were recruited from two tertiary South Korean hospitals, and the age- and sex-matched individuals with typical development were retrospectively collected. We used the AutoMorph pipeline to extract retinal features and used four types of ML models for ADHD screening and EF subdomain prediction, and we adopted the Shapely additive explanation method. ADHD screening models achieved 95.5%-96.9% AUROC. For EF function stratification, the visual and auditory subdomains showed strong (AUROC > 85%) and poor performances, respectively. Our analysis of retinal fundus photographs demonstrated potential as a noninvasive biomarker for ADHD screening and EF deficit stratification in the visual attention domain.

Early Identification of Autism Using Cry Analysis: A Systematic Review and Meta-analysis of Retrospective and Prospective Studies

Cry analysis is emerging as a promising tool for early autism identification. Acoustic features such as fundamental frequency (F0), cry duration, and phonation have shown potential as early vocal biomarkers. This systematic review and meta-analysis aimed to evaluate the diagnostic value of cry characteristics and the role of Machine Learning (ML) in improving autism screening. A comprehensive search of relevant databases was conducted to identify studies examining acoustic cry features in infants with an elevated likelihood of autism. Inclusion criteria focused on retrospective and prospective studies with clear cry feature extraction methods. A meta-analysis was performed to synthesize findings, particularly focusing on differences in F0, and assessing the role of ML-based cry analysis. The review identified eleven studies with consistent acoustic markers, including F0, phonation, duration, amplitude, and voice quality, as reliable indicators of neurodevelopmental differences associated with autism. ML approaches significantly improved screening precision by capturing non-linear patterns in cry data. The meta-analysis of six studies revealed a trend toward higher F0 in autistic infants, although the pooled effect size was not statistically significant. Methodological heterogeneity and small sample sizes were notable limitations across studies. Cry analysis holds promise as a non-invasive, accessible tool for early autism screening, with ML integration enhancing its diagnostic potential. However, the findings emphasize the need for large-scale, longitudinal studies with standardized methodologies to validate its utility and ensure its applicability across diverse populations. Addressing these gaps could establish cry analysis as a cornerstone of early autism identification.

A cross-dataset study on automatic detection of autism spectrum disorder from text data

OBJECTIVE

The goals of this article are as follows. First, to investigate the possibility of detecting autism spectrum disorder (ASD) from text data using the latest generation of machine learning tools. Second, to compare model performance on two datasets of transcribed statements, collected using two different diagnostic tools. Third, to investigate the feasibility of knowledge transfer between models trained on both datasets and check if data augmentation can help alleviate the problem of a small number of observations.

METHOD

We explore two techniques to detect ASD. The first one is based on fine-tuning HerBERT, a BERT-based, monolingual deep transformer neural network. The second one uses the newest, multipurpose text embeddings from OpenAI and a classifier. We apply the methods to two separate datasets of transcribed statements, collected using two different diagnostic tools: thought, language, and communication (TLC) and autism diagnosis observation schedule-2 (ADOS-2). We conducted several cross-dataset experiments in both a zero-shot setting and a setting where models are pretrained on one dataset and then training continues on another to test the possibility of knowledge transfer.

RESULTS

Unlike previous studies, the models we tested obtained average results on ADOS-2 data but reached very good performance of the models in TLC. We did not observe any benefits from knowledge transfer between datasets. We observed relatively poor performance of models trained on augmented data and hypothesize that data augmentation by back translation obfuscates autism-specific signals.

CONCLUSION

The quality of machine learning models that detect ASD from text data is improving, but model results are dependent on the type of input data or diagnostic tool.

A Deep Learning Model to Predict Breast Implant Texture Types Using Ultrasonography Images: Feasibility Development Study

BACKGROUND

Breast implants, including textured variants, have been widely used in aesthetic and reconstructive mammoplasty. However, the textured type, which is one of the shell texture types of breast implants, has been identified as a possible etiologic factor for lymphoma, specifically breast implant-associated anaplastic large cell lymphoma (BIA-ALCL). Identifying the shell texture type of the implant is critical to diagnosing BIA-ALCL. However, distinguishing the shell texture type can be difficult due to the loss of human memory and medical history. An alternative approach is to use ultrasonography, but this method also has limitations in quantitative assessment.

OBJECTIVE

This study aims to determine the feasibility of using a deep learning model to classify the shell texture type of breast implants and make robust predictions from ultrasonography images from heterogeneous sources.

METHODS

A total of 19,502 breast implant images were retrospectively collected from heterogeneous sources, including images captured from both Canon and GE devices, images of ruptured implants, and images without implants, as well as publicly available images. The Canon images were trained using ResNet-50. The model's performance on the Canon dataset was evaluated using stratified 5-fold cross-validation. Additionally, external validation was conducted using the GE and publicly available datasets. The area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (PRAUC) were calculated based on the contribution of the pixels with Gradient-weighted Class Activation Mapping (Grad-CAM). To identify the significant pixels for classification, we masked the pixels that contributed less than 10%, up to a maximum of 100%. To assess the model's robustness to uncertainty, Shannon entropy was calculated for 4 image groups: Canon, GE, ruptured implants, and without implants.

RESULTS

The deep learning model achieved an average AUROC of 0.98 and a PRAUC of 0.88 in the Canon dataset. The model achieved an AUROC of 0.985 and a PRAUC of 0.748 for images captured with GE devices. Additionally, the model predicted an AUROC of 0.909 and a PRAUC of 0.958 for the publicly available dataset. This model maintained the PRAUC values for quantitative validation when masking up to 90% of the least-contributing pixels and the remnant pixels in breast shell layers. Furthermore, the prediction uncertainty increased in the following order: Canon (0.066), GE (0072), ruptured implants (0.371), and no implants (0.777).

CONCLUSIONS

We have demonstrated the feasibility of using deep learning to predict the shell texture type of breast implants. This approach quantifies the shell texture types of breast implants, supporting the first step in the diagnosis of BIA-ALCL.

Development of a deep learning model to distinguish the cause of optic disc atrophy using retinal fundus photography

The differential diagnosis for optic atrophy can be challenging and requires expensive, time-consuming ancillary testing to determine the cause. While Leber's hereditary optic neuropathy (LHON) and optic neuritis (ON) are both clinically significant causes for optic atrophy, both relatively rare in the general population, contributing to limitations in obtaining large imaging datasets. This study therefore aims to develop a deep learning (DL) model based on small datasets that could distinguish the cause of optic disc atrophy using only fundus photography. We retrospectively reviewed fundus photographs of 120 normal eyes, 30 eyes (15 patients) with genetically-confirmed LHON, and 30 eyes (26 patients) with ON. Images were split into a training dataset and a test dataset and used for model training with ResNet-18. To visualize the critical regions in retinal photographs that are highly associated with disease prediction, Gradient-Weighted Class Activation Map (Grad-CAM) was used to generate image-level attention heat maps and to enhance the interpretability of the DL system. In the 3-class classification of normal, LHON, and ON, the area under the receiver operating characteristic curve (AUROC) was 1.0 for normal, 0.988 for LHON, and 0.990 for ON, clearly differentiating each class from the others with an overall total accuracy of 0.93. Specifically, when distinguishing between normal and disease cases, the precision, recall, and F1 scores were perfect at 1.0. Furthermore, in the differentiation of LHON from other conditions, ON from others, and between LHON and ON, we consistently observed precision, recall, and F1 scores of 0.8. The model performance was maintained until only 10% of the pixel values of the image, identified as important by Grad-CAM, were preserved and the rest were masked, followed by retraining and evaluation.