Study Population
In our cohort of 232 TBI patients who received PCC, the median age was 81 years (IQR = [70, 87]) and 54% (125/232) were male. Most were White (59%, 138/232) and of non-Hispanic ethnicity (94%, 218/232), and the predominant mechanism of injury (MOI) was falls (83%, 193/232). At presentation, 81% (187/232) had both pupils reactive and a median mGCS sub-score of 6 (IQR = [5, 6]). During hospitalization, 81% (187/232) of patients were admitted to the ICU with a median ICU length of stay of 6 days (IQR = [3, 13]) and a median TILS score of 2 (IQR = [1, 5]) prior to PCC. Regarding outcomes, 51% (118/232) experienced withdrawal of life-sustaining treatment (WLST), 38% (88/232) experienced inpatient mortality, and an additional 30% (70/232) were discharged to hospice. The median time from admission to PCC was 5 days (IQR = [2, 10]).
Characterization and Comparison of Patient Clusters
Principal component analysis (PCA) captured 8.9%, 5.9%, and 4.9% of the total variance in the first three principal components, respectively. K-means clustering partitioned observations into three distinct clusters, which are visualized using two- and three-dimensional plots (FIGURE 1). Each cluster presented with unique characteristics (TABLE 1). Among the characteristics available at presentation, the clusters differed significantly across 65% (34/52) of all the admission variables recorded (SUPPLEMENTARY TABLE 1).
Among presentation variables, Cluster A was the oldest group (median [IQR] = 87 [78, 94] years) and predominantly consisted of unmarried White females (94% female [83/86], 58% White [51/86], 75% unmarried/widowed [56/86]) with mild TBIs from falls (89%, 78/86) and a high median mGCS sub-score (6 [5, 6]). In comparison, Cluster B was moderately younger (81 [75, 86] years) and comprised mostly married White males (84% male [84/106], 67% white [71/106], 71% married [75/106]), also with mild TBIs from falls (92%, 98/106) and high mGCS sub-score (6 [6, 6]). Finally, Cluster C was much younger (46.5 [29.5, 59.8] years) and included more racial minorities (58% non-White, 22/38) with moderate-severe TBIs, more heterogeneous mechanisms of injury including high-impact MOIs (47%, 22/38), and the lowest median mGCS sub-score (4 [1, 5]). Specifically, Cluster A had the highest percentage of females (p < 0.001); Cluster B had the highest percentage of married patients (p < 0.001) and White patients (p < 0.001); and Cluster C had the most severe injuries with the lowest mGCS sub-scores (p < 0.001) and lowest rate of bilateral reactive pupils at presentation (p < 0.001).
Among hospital course and outcome variables, significant differences were noted in ICU admission rates and TILS scores before PCC, with Cluster C having the highest values (97% ICU admission, 37/38, p = 0.021; median TILS score 5.5 [3.0, 13.8], p < 0.001) (TABLE 2). No other recorded hospital course characteristics differed between clusters. Additionally, none of the clusters differed across rates of WLST, inpatient mortality, or discharge disposition (TABLE 3). However, significant differences were observed in time-to-event data. Cluster A had a shorter median time from admission to PCC (2.5 [1.0, 7.0] days) than Cluster B (5.0 [3.0, 10.8] days, p < 0.001) and Cluster C (9.0 [4.2, 17.0] days, p = 0.034), despite no significant difference in time-to-mortality (TABLE 3; FIGURE 2).
XGBoost Model and SHAP Analysis
An XGBoost model, enhanced by SHAP value analysis for model interpretability, identified age as the most important factor influencing time-to-PCC, followed by male sex and White race. Notably, these demographic factors were more influential than clinical factors, such as neurological presentation, radiographic findings, comorbidities, mechanism of injury, and care-limiting directive present on admission (FIGURE 3).
The SHAP analysis offered a visual and comprehensive summary of the impact of each feature on the model (SUPPLEMENTARY FIGURE 2). Further investigation revealed nonlinear relationships between age, race, sex, and BMI, identified through interaction evaluations using a SHAP heatmap (SUPPLEMENTARY FIGURE 3) and scatterplots (FIGURE 4). Specifically, the strongest interaction indicated that younger patients with lower BMI were more likely to experience longer time-to-PCC, while the opposite was true for patients with low BMI older than 70 years (FIGURE 4). Other important interactions included Age by Race and Age by Sex, suggesting that younger non-White and younger female patients were also more likely to experience longer time-to-PCC. As with the association between age and BMI, the associations of age with race and sex also reversed directionality for patients older than 70 years (FIGURE 4).