Abstracts

Rationale: Diverse language outcomes have been observed among children with drug-resistant epilepsy (DRE) originating from the left (language-dominant) hemisphere, typically being categorized into one of the following subtypes: no appreciable impact (NI, intact verbal and non-verbal skills), typical organization (TO, verbal skill impaired with the preservation of non-verbal skill), “crowding” effect (CE, non-verbal skill impairment with the preservation of verbal skill, an observation attributable to reorganization of the verbal function to the right hemisphere) and global impact (GI, severely impaired verbal and non-verbal skills). To guide targeted therapeutic interventions, this study presents a novel deep learning approach that can subtype diverse language outcomes: NI, TO, CE, GI, from clinical DWI connectome (DWIC) data of DRE children with left hemispheric epilepsy.

Methods: 24 DRE children (mean age:12.0±3.9 years) with left hemispheric lesions underwent preoperative DWI tractography scans and neuropsychological assessments (verbal IQ, non-verbal IQ, and clinical evaluation of language fundamentals) to identify the ground-truth targets of four subtypes, NI (n=6, verbal and non-verbal IQ ≥ 87), TO (n=7, verbal IQ < 87, non-verbal IQ ≥ 87), CE (n=5, verbal IQ ≥ 87, non-verbal IQ < 87), and GI (n=6, verbal IQ < 87, non-verbal IQ < 87). Our previous tract classification method was applied to construct whole brain backbone network DWIC, S(k,l) in which elements define average fractional anisotropy values of 1477 pair-wise pathways connecting k^th and l^th cortical parcellations. Verbal network, V(m,n) and non-verbal network, NV(m,n) were constructed by gathering all elements of S(k,l) of which values are signficantly correlated with verbal and non-verbal IQ scores across DRE patients. Data from 29 healthy controls (mean age: 11.6±3.3 years) provided normative magnitudes of V(m,n) and NV(m,n). A deep ensemble learning network (Fig. 1) was implemented to predict the ground-truth clinical subtypes using a set of input features: V(m,n), NV(m,n), and clinical variables.

Results: Significant subtype differences were found in the node-average Z-scores of local efficiency: Z(LE) within V(m,n) and NV(m,n), that quantifies how much the efficiency of neighboring connectivity deviates from that of age-matched controls (Fig. 2A). Compared with the controls, GI showed significantly lower values of the averaged Z(LE) in V(m,n) and NV(m,n). Also, the reduction in V(m,n) was more significant in both GI and TO. Compared with the TO, the CE showed significantly higher values of the average Z(LE) in V(m,n) of the left hemisphere and significantly lower value of the average Z(LE) in NV(m,n) of the right hemisphere, providing an imaging correlate of the crowding effect. A high accuracy (0.91±0.02) was obtained when the four subtypes were predicted in 500 leave-one-out cross validation trials (Fig. 2B).

Conclusions: This study presents a new imaging tool capable of identifying diverse subtypes of preoperative language and non-verbal skills, which may potentially guide targeted therapeutic interventions for children with drug-resistant left-hemispheric epilepsy.

Funding: NIH R01NS089659, R01NS064033, and R01 NS041922.