Improving health-related quality of life (QoL) is a primary goal in managing Parkinson's disease (PD). Deep brain stimulation (DBS) of the subthalamic nucleus (STN) can improve motor function, but QoL outcomes vary considerably, with up to half of patients reporting no meaningful improvement. Current approaches lack integration of demographic, patient-reported, neuroimaging, and neurophysiological data to explain this variability. This complexity hinders optimal therapeutic recommendations. While preoperative baseline QoL is often cited as a predictor, with higher initial QoL burden associated with greater gains, contradictory findings exist. Furthermore, electrode location and stimulation-induced neural responses also impact outcomes, highlighting the need for a multimodal approach integrating patient-reported data, neuroimaging, and neurophysiological data to understand and predict QoL changes after STN DBS. This study aimed to address this gap by using explainable machine learning to elucidate the response variability of STN DBS in terms of QoL, considering preoperative patient data, intraoperative neurophysiological metrics, and perioperative neuroimaging findings.

Existing literature shows inconsistent results regarding predictors of postoperative QoL improvement after STN DBS. While some studies consistently identify preoperative baseline QoL as the strongest predictor, with greater QoL burden associated with greater postoperative gains, others report an inverse relationship or the influence of other factors. Research has also explored the role of surgical factors, such as variations in electrode contact location within the STN, and the magnitude of stimulation-induced neural responses. While the influence of electrode placement on motor symptoms is well-documented, its impact on QoL variability is less clear. Studies investigating the integration of patient-reported, neuroimaging, and neurophysiological data to explain QoL variability are sparse, highlighting a critical knowledge gap. The application of machine learning in PD research primarily focuses on motor symptoms, often overlooking QoL.

This study employed explainable machine learning using an XGBoost Regressor model with leave-one-out cross-validation to predict changes in QoL, measured by the Parkinson's Disease Questionnaire (PDQ-39), in 63 PD patients who underwent STN DBS. The analysis incorporated demographic information, preoperative PDQ-39 scores, electrode positions, and intraoperative neural recordings along the implantation trajectory. SHAP (SHapley Additive exPlanations) values were used to interpret the contribution of each variable to the model's predictions. Electrophysiological data were collected at multiple depths during implantation, encompassing theta, alpha, lower beta, and upper beta frequency bands. Electrode positions were determined in standard MNI space. The model predicted changes in PDQ-39 scores, normalized to a range between -1 and 1, with scores closer to 1 indicating improvement and those closer to -1 indicating deterioration. Model performance was evaluated using Pearson's correlation coefficient and mean squared error. SHAP values were used for feature importance analysis, and an ablation analysis was performed to assess the individual contribution of each feature. A support vector machine (SVM) was employed to determine thresholds for each feature that predicted positive contributions to the model.

The study found a statistically significant correlation between preoperative PDQ-39 scores and normalized PDQ-39 changes (p=3.15e-3, r=0.36). Patients with higher preoperative QoL burden experienced greater postoperative improvements. Incorporating additional features, including electrophysiological data, improved model performance. The electrophysiological data, particularly from the left STN, showed high predictive capability, regardless of the chosen approach (along implantation trajectory, averaged over contacts, or at active contacts). SHAP analysis identified preoperative PDQ-39 score and upper beta band activity (>20 Hz) in the left STN as the most important predictors of QoL changes. Higher preoperative PDQ-39 scores were associated with greater improvements, while higher upper beta activity in the left STN predicted improvement. Postoperative levodopa reduction (LEDD ratio), age, and time since surgery also influenced the predictions. Electrode position along the superior-inferior axis (z-coordinate), specifically relative to z = -7 in standard MNI space, significantly influenced QoL, with positions above this coordinate associated with improvement and those below with worsening. Ablation analysis confirmed the importance of preoperative PDQ-39 and left STN upper beta activity for model performance. Replacing upper beta activity with baseline UPDRS-3 score significantly reduced model performance, suggesting that upper beta activity is a unique predictor independent of motor symptom severity. The SVM analysis yielded thresholds for each feature that accurately predicted improvement: PDQ-39 > 31.5 points, upper beta band activity > 0.15 V²/Hz in left STN, LEDD ratio < 0.68, and age < 71.5 years.

This study addresses the significant variability in QoL outcomes following STN DBS for PD by integrating multimodal data and employing explainable machine learning. The findings highlight the crucial role of both baseline patient-reported QoL and precise electrode placement within the STN, particularly emphasizing the importance of upper beta activity in the left STN. The strong predictive value of preoperative QoL confirms prior research, demonstrating that patients with higher initial QoL burden tend to experience greater improvements. The identification of upper beta activity as a key predictor extends previous research, suggesting its potential as a biomarker for QoL improvement independent of motor symptom severity. The influence of electrode location along the dorso-ventral axis adds another dimension, emphasizing the need for precise targeting for optimal QoL outcomes. The study's findings contribute significantly to the understanding of factors influencing QoL in STN DBS, offering potential for improved patient selection, surgical targeting, and stimulation programming strategies. The results suggest the need for a more holistic approach to STN DBS, moving beyond a focus on motor symptoms to encompass the crucial aspect of QoL.

This study demonstrates that integrating multimodal data and employing explainable machine learning can significantly improve the prediction of QoL outcomes after STN DBS for PD. Preoperative PDQ-39 scores and upper beta activity in the left STN are key predictors, independent of motor symptom severity. Electrode placement, particularly along the z-axis, also significantly impacts QoL. These findings highlight the need for personalized approaches to DBS, emphasizing the integration of patient-reported data and neurophysiological measures for optimizing treatment and improving QoL. Future research should focus on validating these findings in larger cohorts and exploring the precise mechanisms linking upper beta activity and QoL.

The study's sample size is relatively small, potentially limiting the generalizability of the findings. The study focused on a specific DBS system and patient population, and the results may not be directly transferable to other settings. While the study used SHAP values to explain the model, the complex interplay between features might still be difficult to fully interpret. Future studies with larger, more diverse cohorts are needed to validate and extend these findings.