In a recent study published in npj Mental Health Research, researchers performed a systematic review to determine whether artificial intelligence (AI) may be a promising strategy for post-traumatic stress disorder (PTSD) diagnosis.
Study: Systematic review of machine learning in PTSD studies for automated diagnosis evaluation. Image Credit: Suriyawut Suriya/Shutterstock.com
Background
Machine learning (ML) approaches have been used to diagnose and treat PTSD, an illness frequently misdiagnosed because of its complex clinical and biochemical aspects.
This method can potentially improve clinical results while also lowering the costs associated with long-term handicaps, lowering the burden on individuals globally.
About the study
In the present study, researchers presented the role of machine learning in automated PTSD diagnosis.
Statistical approaches were utilized to aggregate the findings of included studies and give advice on crucial machine learning task implementation aspects.
These included (i) selecting the most suitable machine learning model for the dataset, (ii) identifying optimal machine learning features based on the selected method for diagnosis, (iii) determining an optimal sample size according to data distribution, and (iv) implementing appropriate tools to evaluate and validate model performance.
Databases such as Embase, MEDLINE, Scopus, PsycINFO, Compendex, and IEEE Xplore, spanning from 1946 to 2022, were searched through 18 October 2022. Two researchers independently screened the data and resolved discrepancies by discussion.
Only records published in English that employed machine learning-based models to conduct automated PTSD diagnosis and specified assessment measures used to report model performance were included.
Duplicate records, case studies, non-journal articles, records unrelated to PTSD diagnosis, records with no mention of evaluation metrics, records primarily concerned with feature selection, and records focusing on the prediction of future outcomes, risk factors for acquiring PTSD, or the trajectory of symptoms rather than PTSD diagnosis were excluded from the analysis.
The current study's quality was determined using the provided evaluation metrics [such as accuracy, precision, recall, F1 scores, and area under the curve (AUC) values] from the selected studies.
Results
A total of 3,186 records were identified, of which 1,654 remained after duplicate removal. After title and abstract screening, 1,502 records were excluded, and therefore, 152 studies underwent full-text screening, of which 111 records that did not satisfy the eligibility criteria were excluded. As a result, 41 records were considered for the final analysis.
The sample size ranged from 24 individuals to 2,124,496 surveys, with a median of 179 samples across the analyzed research. The K-fold cross-validation (30 studies) and Support Vector Machine (SVM, 12 studies) models were most widely used in the included studies. Deep learning (DL), SVM, and mixed models outperformed others in the included studies.
Support Vector Machine, a long-standing machine learning model, was selected because of its high efficiency and adaptability for small and moderate-sized data and its comparatively modest processing requirements.
Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) AI models, including Long Short-Term Memory (LSTM), were used as classifiers. The amygdala, hippocampus, prefrontal cortex, and insula provided significant predictive findings when obtained using functional magnetic resonance imaging (fMRI) techniques.
ML models were increasingly used to diagnose PTSD, particularly in clinical settings where the chance of PTSD symptoms is high. However, the high expenses of specialized equipment and the need for experienced personnel limit the use of these models.
DL models required less feature engineering and were particularly beneficial for PTSD classification due to their capacity to choose informative features.
In contrast, research relying on self-documented questionnaires and other approaches for PTSD diagnosis, such as online databases or surveys, typically had higher sample numbers. These data sources were less PTSD-specific, with varied samples, making it difficult to derive effective predictive characteristics for PTSD screening.
Data imbalance could also lead to overfitting the machine learning models and loss of generalizability of the results. Limited sample sizes, comorbidities, inadequate study controls, generalizability loss, and uneven data distribution influenced the efficacy of ML models for PTSD classification.
Overfitting as a result of small sample numbers could reduce model precision, recall, and accuracy. Minority communities may be underrepresented as a result of data imbalances, and ethical and legal issues must be highlighted.
Data appropriateness and computational resources were important factors in determining optimal ML models. Traditional ML models, such as SVM and ensemble models, were preferably used for data significantly linked with PTSD diagnosis and needing minimum feature engineering, such as neuroimaging data.
DL models, or MLPs, were considered acceptable for complex audio, visual, and textual data. Data imbalances may be addressed via resampling and ensemble approaches. Model validation is critical for ensuring ML correctness and dependability.
By splitting the data and training the model on certain partitions while assessing others, cross-validation could minimize model variance. Permutation tests must be carried out to verify model dependability and eliminate stochastic effects.
Overall, the study findings highlighted the potential of AI in diagnosing post-traumatic stress disorder. AI can provide cost-effective, dependable, and quick approaches for diagnosing PTSD, particularly for stigmatized people who have trouble getting appropriate mental healthcare.
However, due to ethical and privacy concerns, as well as a lack of established rules, the actual application of AI systems still requires refinement. The findings may be used to drive future research in automated PTSD diagnosis, emphasizing the importance of AI in early PTSD diagnosis.
Physical activity lowers cardiovascular disease risk by reducing stress-related brain activity