Mental health significantly impacts the quality of life for cancer patients, often leading to severe consequences such as depression, anxiety, and even suicide. Current methods of mental health evaluation rely heavily on subjective questionnaires, lacking objectivity and efficiency. This research addresses this gap by developing an objective method using visual metrics, specifically eye movements, recorded via TobiiPro eyeglasses. Eye movements are closely linked to brain activity and have shown correlations with various cognitive and mental disorders. Deep learning, particularly CNN-LSTM models, are well-suited for analyzing time-series data like eye-tracking information and identifying patterns indicative of mental health states. The study aims to leverage this technology to create an automated, objective assessment tool for mental health in cancer patients post-surgery, improving patient care and streamlining access to mental health services. Existing objective methods are limited, often requiring feature engineering and lacking integration with clinically approved assessments. This study bridges this gap by utilizing a clinically validated assessment method alongside advanced deep learning techniques.

The literature highlights the strong association between mental health issues and cancer, with high rates of depression and anxiety reported among cancer patients. The existing methods of mental health evaluation are predominantly subjective, relying on self-reported questionnaires. While some studies have explored objective methods using machine learning, they often lack integration with clinically validated assessments and require significant feature engineering. The study reviews existing research in objective mental health evaluation, pointing out their limitations. It also examines the existing body of research that has established a relationship between ocular motor functions and various cognitive and mental disorders. The use of eye movement analysis as a potential biomarker for mental health is established as the foundation for the proposed deep learning methodology.

The study involved 25 participants (16 cancer patients post-major oncologic surgery and 9 healthy controls). Participants viewed 18 artworks in a controlled gallery setting while their eye movements were recorded using TobiiPro eyeglasses at 100 Hz. The recorded data included 20 visual metrics, including gaze point coordinates, gaze direction, and pupil diameter. Pre- and post-study questionnaires assessed hope (HHI), anxiety (STAI), and mental well-being (WEMWBS). Clinical psychotherapy and statistical methods defined cutoff scores to categorize participants into low, intermediate, and high levels for each mental health metric. A CNN-LSTM deep neural network was trained on the visual time-series data to classify individuals into these three levels for each metric. The CNN component of the model extracted features from the visual data, while the LSTM component processed the temporal dependencies inherent in the eye movement time series. The moving average filter was used for noise reduction in the gaze data. The model's architecture is comprised of 1D convolutional layers, dropout layers, max-pooling layers, flattening layers, an LSTM layer, fully connected layers, and finally a softmax activation function at the output layer for three-class classification. The model's performance was evaluated using classification accuracy.

The developed CNN-LSTM model demonstrated high accuracy in classifying mental health metrics based on visual data. The classification accuracy was 93.81% for the HHI (hope), 94.76% for the STAI (anxiety), and 95.00% for the WEMWBS (mental well-being). These high accuracies suggest the model's potential for objective mental health assessment using easily collected visual data. The model effectively learned patterns in eye movement data that reflect different levels of hope, anxiety, and mental well-being. This demonstrates the potential to use visual data as a proxy for subjective mental health assessments.

The high accuracy achieved by the CNN-LSTM model demonstrates the feasibility of using visual metrics as an objective measure of mental health in cancer patients and controls. The results support the hypothesis that patterns in eye movements provide valuable information regarding mental state. The study shows that deep learning can successfully be applied to automate the assessment process, eliminating the need for subjective self-report measures. This objective assessment can significantly improve the efficiency and reliability of mental health evaluations in clinical practice and could allow for more timely and effective interventions. The findings contribute to the growing body of research exploring the use of technology to improve mental health care, particularly among vulnerable populations like cancer patients.

This study successfully developed and validated a deep learning model for objective mental health assessment using visual metrics obtained from eye-tracking technology. The high accuracy of the model suggests its potential for real-world applications in home-based mental health monitoring, particularly for cancer patients post-surgery. Future research could focus on larger datasets, expanding the range of mental health conditions evaluated, and integrating the model into mobile applications for broader accessibility and usability. Further investigations into the specific features extracted by the CNN are needed to improve the interpretability of the model and deepen our understanding of the neurophysiological processes behind the relationship between eye movements and mental health.

The study's relatively small sample size might limit the generalizability of the findings. The study population primarily comprised patients undergoing specific oncologic procedures. The generalizability of the model across diverse populations and cancer types warrants further investigation. Additionally, the ecological validity of the art gallery setting should be considered when applying the model to real-world scenarios. Future research could incorporate more diverse stimuli and environments.