Non-Invasive Mapping Predicts Language Outcomes after Eloquent Tumor Resection

108 

Abstract Submission 

Matthew Muir¹, Kyle Noll¹, Hayley Michener¹, Sarah Prinsloo¹, Sujit Prabhu¹

¹MD Anderson Cancer Center, Houston, TX

Matthew Muir  
MD Anderson Cancer Center
Houston, TX

Kyle Noll, PhD  
MD Anderson Cancer Center
Houston, TX

Hayley Michener, MS  
MD Anderson Cancer Center
Houston, TX

Sarah Prinsloo, PhD  
MD Anderson Cancer Center
Houston, TX

Sujit Prabhu, MD  
MD Anderson Cancer Center
Houston, TX

Glioma patients undergoing surgery in eloquent regions consistently sustain permanent postoperative language deficits that decrease both quality of life and survival. The origins of these poor outcomes remain unknown. Despite the advent of intraoperative mapping techniques, subjective judgements frequently determine important surgical decisions. Transcranial magnetic stimulation (TMS) has recently emerged as a promising non-invasive, preoperative language mapping technique. We aim to elucidate the determinants of aphasic surgical deficits by building an individualized predictive model based on TMS, routinely acquired preoperative imaging data, and the resection volume. The results shed light on the structure and function of large-scale language networks in glioma patients and lead to a clinical imaging approach for predicting and avoiding postoperative aphasic decline.

This retrospective study included 79 consecutive patients who underwent preoperative TMS language mapping and subsequent awake craniotomy for the resection of language eloquent gliomas. We used a deformable registration algorithm to co-register the postoperative MRI with the preoperative MRI containing functional and structural imaging features. We correlated the resection versus preservation of regions identified by preoperative reconstructions with pre to postoperative changes in the Western Aphasia Battery. We used TMS points as a collective seed for fiber tracking. We used a fractional anisotropic threshold selection approach standardized to the individual profile of each patient (25%, 50%, 75%, and 85%). We normalized the resected portion of the tracts to MNI space and analyzed their relationship with normative white matter tracts (7 language associated tracts: AF, SLF, IFOF, ILF, MdLF, FAT, UF) from the Human Connectome Project. We used binary logistic regression and confusion matrix elements to evaluate the predictive value of each model. We determined 1 versus 0 predictions if the region identified by the cortical or subcortical reconstruction was resected or not. We determined 1 versus 0 outcomes based on the language status of the patient at 1-2 months postoperatively (aphasia was counted as a 1).

While the resection of TMS points alone did not significantly predict postoperative outcome (OR=2.8, p=.15), the resection of TMS points with robust subcortical connectivity with different fractional anisotropic profiles significantly predicted aphasic deficits at every threshold (Figure 1B: 25%-OR=6.1, p=.011; 50%- OR=5.1, p=.017; 75%- OR=8.6, p=.004; 85%- OR=5.9, p=.019). An ROC curve based on the FA threshold for different connectivity groups of TMS points showed an AUC of .72 (Figure 1A). The same ROC curve for the subcortical TMS seeded tracts showed an AUC of .89 (Figure 1C). Multivariate analysis revealed that the resection of subcortical tracts independently predicts aphasic deficits while cortical resections do not (Figure 1D). We found that true positive tract-resection volumes significantly associated with normative HCP tracts compared to false positive tract-resection volumes (Figure 2). We used these measures to preoperatively predict the functional subdomains of TMS tracts based on their interaction with HCP volumes. This improved the positive predictive value (PPV) of the resulting composite model (TMS + tracts + HCP classification) from 50% (TMS + tracts) to 89%. This final model showed an odds ratio of 155 (p<.001), PPV of 89%, negative predictive value of 95%, sensitivity of 84%, a specificity of 97%, and an accuracy of 94%.

These results show that non-compensable language function in glioma patients localizes at the subcortical level to individualized subdomains within the normative structure of canonical white matter tracts. We integrate the findings into an actionable preoperative predictive model for permanent aphasic surgical deficits based on the resection versus preservation of regions identified by non-invasive imaging data.

TMS ¹

Language Acquisition

Language Comprehension and Semantics

Language Other

White Matter Anatomy, Fiber Pathways and Connectivity ²

Language

Plasticity

Spatial Normalization

Sub-Cortical

Tractography

Transcranial Magnetic Stimulation (TMS)

White Matter

WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC

Other - Gliomas; Tumor resection