Predicting panic disorder with machine learning of resting-state fMRI features

523 

Abstract Submission 

Junhyung Kim¹, Byung‑Hoon Kim²

¹Department of Psychiatry,KoreaUniversity College of Medicine, Seoul, Republic of Korea, ²Department of Psychiatry,Yonsei University College of Medicine, Seoul, Republic of Korea

Junhyung Kim  
Department of Psychiatry,KoreaUniversity College of Medicine
Seoul, Republic of Korea

Byung‑Hoon Kim  
Department of Psychiatry,Yonsei University College of Medicine
Seoul, Republic of Korea

Panic Disorder (PD) is a condition marked by unexpected, intense panic attacks, often in minimally threatening situations.1 PD can significantly disrupt daily life. While neuroimaging has identified PD-related brain regions, using these findings to predict or classify PD remains a relatively unexplored area. This study employs radiomics to classify PD using fMRI-derived features like regional homogeneity (ReHo) and fractional amplitude of low-frequency fluctuation (fALFF).2 These features, reflecting brain activity and connectivity, are integrated with machine learning models to determine PD diagnosis. The aim is to evaluate the effectiveness of these models in distinguishing PD patients from healthy controls, focusing on the relative importance of various fMRI-derived features. This approach could deepen our understanding of PD's neurobiological aspects in young adults.

This study at Korea University Guro Hospital involved 29 patients with PD and 28 healthy adults, categorized into Panic and CON groups. Participants underwent psychiatric evaluations, including the Mini-International Neuropsychiatric Interview, to screen for psychiatric illnesses and drug use. The study, ethically approved (2021GR0321), utilized various scales like the Panic Disorder Severity Scale (PDSS) and Hospital Anxiety and Depression Scale (HADS) to assess anxiety severity and traits. Imaging data were collected using a 3 T Philips Ingenia scanner, capturing coronal anatomical and functional images. Data preprocessing focused on resting-state brain functional radiomic features like ReHo and fALFF. These features were extracted for 14 brain regions of interest (ROIs) related to anxiety disorders. The dataset was processed using Python libraries pandas and scikit-learn, then split into training and test sets, with the former normalized for stable training. Six machine learning models were employed, including Logistic Regression, Support Vector Machine, Random Forest, and eXtreme Gradient Boosting (XGBoost). Model optimization used GridSearchCV for hyperparameter tuning, and performance was evaluated using the F1 score. The study's final evaluation of the test dataset involved measuring accuracy and F1 score, along with the average and standard deviation of predicted class probabilities.

The distributions of age, sex, and marital status between the two groups did not vary significantly. The PD group had more years of education than the CON group. Except for the LSAS-fear and LSAS-avoidance subscale scores, psychological characteristics associated with anxiety differed significantly. The accuracy of the logistic regression (LogReg), SVM, random forest (RF), multi-layer perceptron (MLP), and the extreme gradient boosting (XGBoost) for classifying the PD group was 0.315, 0.315, 0.515, 0.616, and 0.616, respectively. The F1 score, which reflects both the sensitivity and specificity of the model by computing the harmonic mean of the precision and recall scores, also resulted in a similar trend with the accuracy and the balanced accuracy, demonstrating 0.530, 0.530, 0.626, 0.714 and 0.714 for the LogReg, SVM, RF, MLP, and XGBoost models, respectively.

It could be seen that the XGBoost model resulted in the best classification performance on the test data in terms of all three metrics evaluated.

Psychiatric (eg. Depression, Anxiety, Schizophrenia) ¹

Task-Independent and Resting-State Analysis ²

Anxiety

Limbic Systems

Machine Learning

MRI