Prediction of Headache Improvement Using Multimodal Machine Learning in Patients with Acute Post-traumatic Headache
Amogh Joshi1, Md Mahfuzur Rahman Siddiquee1, Jay Shah1, Todd Schwedt2, Catherine Chong2, Baoxin Li3, Teresa Wu1
1Arizona State University, 2Mayo Clinic, 3Arizona State University, School of Computing and Augmented Intelligence
Objective:
To develop a machine learning method on multimodal data for predicting headache improvement in patients with acute post-traumatic headache (aPTH) attributed to mild traumatic brain injury.
Background:
To assess headache improvement in aPTH patients, we used multimodal data including brain MRI T2* imaging and resting-state functional connectivity among 63 pain-processing areas. Speech samples over 12 weeks were collected. Patients are deemed "improved" if post-injury excess headaches in the third month decrease compared to the first month or are under 2.5; otherwise, they are labeled "not-improved."
Design/Methods:
We analyzed 43 aPTH patients (27 females/16 males) within 0-59 days post-mild TBI, comparing them to 61 healthy controls (39 females/22 males). Our multimodal approach combined T2* features and speech data with an SVM classifier. Key predictive features for headache improvement were selected using Scikit-learn's SelectKBest method. Additionally, we explored combining functional connectivity and T2* data using a Graph Neural Network (GNN) and ensembled it with an SVM trained on speech data.
Results:
At the 3-month follow-up, 26 aPTH patients saw headache improvement, while 17 did not. SVM classifier for predicting headache improvement for single-modality data revealed an AUC of 0.658 using T2* features of five brain regions (right posterior insula, bilateral somatomotor, right spinal trigeminal nucleus, and right middle frontal) and an AUC of 0.749 using speech features including ratio of demonstrative words, jitter relative average perturbation, pitch perturbation factor, and recurrence period density entropy. Combining both modalities improved performance to an AUC of 0.866. A GNN combining T2* and fMRI data with speech data resulted in a lower AUC of 0.60, likely due to limited sample size and increased input dimensionality, affecting the GNN's ability to learn effective features.
Conclusions:
PTH improvement at three months is accurately predicted by SVM on T2* and speech data.