Summary: New AI technology can identify the type of post-stroke depression in patients, enabling more targeted treatment and potentially improving recovery.
Source: Hiroshima University
Researchers at Hiroshima University have developed an artificial intelligence system that can help clinicians detect different types of post-stroke mood disorders—such as depression, apathy, and anxiety—by analyzing routine clinical evaluation data. Because mood disorders after stroke are common but frequently underdiagnosed, this tool could support earlier, more accurate treatment decisions and improve rehabilitation outcomes.
The team trained a probabilistic artificial neural network known as a log-linearized Gaussian mixture network to classify mood disorder types. The model used 36 evaluation indices drawn from functional, physical, and cognitive tests administered to 274 stroke patients. These indices reflect everyday activities, degree of paralysis, stress awareness, and higher brain functions.
The research, which examines the links between motor and cognitive function and subsequent post-stroke mood disorders using machine learning, is reported in the journal Scientific Reports.
Early detection of post-stroke depression (PSD)
The researchers emphasize that different PSD types may arise from distinct neuroanatomic and psychological mechanisms and that each type can affect recovery in unique ways. Detecting these mood disorders early is therefore essential to provide the right therapeutic approach at the right time.
“Depression commonly occurs in the acute and subacute phases after stroke and is known to negatively affect both functional and cognitive recovery,” says Seiji Hama, a research associate at the Graduate School of Biomedical and Health Sciences, Hiroshima University, and a co-author of the study. “Early diagnosis and intervention are crucial for post-stroke depression.”
Hama and colleagues note that PSD is multifactorial and that overlapping neurological symptoms often obscure mood disorder detection. Their study demonstrates that routinely collected clinical data—without special equipment—can be used to support accurate PSD diagnosis, marking a practical first step toward wider clinical application.
To evaluate the model’s diagnostic performance, the team used the receiver operating characteristic (ROC) curve and measured the area under the curve (AUC). The AI achieved AUC scores above 0.85, indicating moderate to high discrimination ability for identifying the three mood disorder types from the provided clinical indices.
Stress threshold hypothesis and mechanisms
Post-stroke physical impairments, cognitive dysfunction, and mood disturbances are deeply interconnected and often involve stress-related responses, which makes it challenging to pinpoint the causes of PSD. It remains unclear in many cases whether mood changes reflect a psychological response to loss of function or biological effects of brain injury.
The study’s findings support a “stress threshold hypothesis,” suggesting that brain lesions caused by stroke may reduce a patient’s capacity to adapt to stress and thereby increase vulnerability to mood disorders. This perspective aligns with earlier reports linking accumulated small infarcts in regions such as the basal ganglia, thalamus, and deep white matter with higher PSD risk.
The authors plan further research to clarify the neural origins of PSD by integrating MRI-based analyses with their machine learning approach. They hope that combining imaging findings with routine clinical indices will refine diagnostic accuracy and reveal specific lesion patterns linked to different mood disorder types.
Hama also noted the longer-term goal of adapting this diagnostic technique for wearable or community-based devices. If feasible, such tools could enable ongoing monitoring of cognitive and emotional status after stroke, support early intervention in outpatient or home settings, and potentially contribute to stroke prevention strategies by detecting cognitive decline earlier.
About this AI research news
NeuroscienceNews would like to thank Mikas Matsuzawa for submitting this AI research news.
Source: Hiroshima University
Contact: Mikas Matsuzawa – Hiroshima University
Image: The image is in the public domain
Original Research: Open access.
“Relationships between motor and cognitive functions and subsequent post-stroke mood disorders revealed by machine learning analysis” by Seiji Hama, Kazumasa Yoshimura, Akiko Yanagawa, Koji Shimonaga, Akira Furui, Zu Soh, Shinya Nishino, Harutoyo Hirano, Shigeto Yamawaki & Toshio Tsuji. Published in Scientific Reports.
Abstract
Relationships between motor and cognitive functions and subsequent post-stroke mood disorders revealed by machine learning analysis
Mood disorders—including depression, apathy, and anxiety—are common after stroke and can hinder functional recovery, interacting with various physical and cognitive impairments. These overlapping symptoms make post-stroke mood disorders complex and difficult to interpret.
This study sought to clarify cross-sectional relationships between mood disorders and motor/cognitive functions in stroke patients. The authors designed an artificial neural network to predict three mood disorder types from 36 clinical evaluation indices collected from 274 patients.
By applying input dimensionality reduction, the analysis identified which evaluation indices are most strongly related to each mood disorder. The model’s ROC analysis showed AUC values above 0.85, indicating solid predictive performance. The results support a stress threshold hypothesis in which stroke-induced brain lesions increase stress vulnerability and may precipitate mood disorders.