Comorbidity Characterization
We included 2,629 patients in our analysis. Mean age was 59.0, 61.9% were male, average GCS was 10.4, the mean number of comorbidities was 2.0, and the rate of survival to discharge was 36.7%. Demographics of study cohort are summarized in Table 1.
Table 1
Study demographics of 2,629 included patients Parameters of age, sex, Glasgow coma scale (GCS), number of comorbidities, survival rate, intervention rate, and hospital length of stay for the traumatic brain injury cohort. Continuous variables reported as mean ± SD. Categorical variables reported as count with percentage of population.
Population parameter | Value |
Mean age | 59 ± 23 years |
Male sex | 1,627 (61.9%) |
Mean (Median) GCS | 10.4 (11) ± 4.6 |
Mean (Median) number of comorbidities | 2.1 (2) ± 1.8 |
Survival to discharge | 1665 (63.3%) |
Neurosurgical intervention | 604 (23.0%) |
Mean length of stay | 9.3 days ± 10.9 days |
The relationship between age, number of comorbidities, and survival can be visualized in Fig. 1. Age and comorbidity count was only weakly associated (r = 0.20, p < 0.001). We then tested the association of the comorbidity count with the survival to discharge after accounting for age and GCS score in a multivariable logistic regression model and found that the number of comorbidities was negatively associated with survival (OR for survival per one increase in comorbidity count = 0.984, CI: 0.975–0.994).
We next investigated the pattern of comorbidities, specifically their associations in the entire cohort to test our hypothesis that many comorbidities co-occur. These relationships between the Elixhauser comorbidity groups are visualized by the network graph in Fig. 2. This network discovery of the data revealed intuitive subgroups: congestive heart failure (CHF) with arrhythmia (RR = 2.76, CI: 2.73–2.79), drug abuse with alcohol abuse (RR = 3.02, CI: 2.95–3.11), and renal failure with complicated hypertension (RR = 16.27, CI: 15.96–16.58).
Clinical Characterization of Comorbidity Endotypes
We next analyzed the comorbidity data for the presence of endotypes using LCA. Five stable clusters of patient endotypes resulted, detailed in Fig. 3 by their comorbidity probability distribution. Endotype 1 was labeled Heart Failure and Arrhythmia (HFA), characterized by cardiac comorbidities. Endotype 2 featured few comorbidities and was labelled Healthy (HE) and used as the reference endotype for subsequent analyses. Endotype 3 is characterized mainly by high rates of renal failure with complicated hypertension and was labelled Renal Failure with Hypertension (RFH). Endotype 4 was labelled Alcohol Abuse (AA), featuring high rates of alcohol abuse and relatively low rates of other comorbidities. Endotype 5 had high rates of hypertension and relatively low rates of other comorbidities, labelled Hypertension (HTN).
The distribution of TBI injury severity categories was relatively similar although HE and AA endotypes contained slightly larger proportions of severe TBI at 46.9% and 39.4%, respectively (Table 2). The HFA and RFH endotypes had nearly identical age distributions of mostly older individuals in addition to a small subpopulation of middle-aged individuals (17.9% middle-aged, 82.4% old vs 19.7% middle-aged, 80.9% old, respectively). The HTN endotype was composed mainly of older adults (68.6%) with a moderate proportion of middle-aged individuals (28.3%) and a small subpopulation of young individuals (3.7%). The AA group was primarily composed of middle-aged individuals (73.7%). HE endotype was composed of mainly young (47.6%) and middle-aged (40.0%) individuals. There were minor sex differences among a few endotypes, contributed to mainly by the young age segment which consistently had a higher proportion of male patients; middle-aged and old age groups across endotypes had comparable sex distributions (Supplementary Table 2). For comorbidity count, HE had the least number of comorbidities (0.8 ± 0.8) followed by HTN (2.5 ± 1.2) while the other comorbidities had higher averages within a similar range (AA: 4.0 ± 1.2, HFA: 4.5 ± 1.6, RFH: 5.5 ± 2.1).
Table 2
Demographics of endotypes Average age, number of comorbidities, and breakdown of injury severity by Glasgow coma scale (GCS) and age categories across the five stable comorbidity endotypes. HFA: Heart failure and arrythmia, HE: Healthy, RFH: Renal failure with hypertension, AA: Alcohol abuse, HTN: Hypertension. Young: 16–39, Middle-aged: 40–69, Old: 70+. Mild: GCS 14–15, Moderate: GCS 9–13, Severe: GCS 3–8.
Grouping | HFA | HE | RFH | AA | HTN |
Mild | 92 (50.0%) | 454 (36.9%) | 71 (53.8%) | 71 (40.6%) | 443 (51.6%) |
Moderate | 36 (19.6%) | 200 (16.2%) | 23 (17.4%) | 35 (20.0%) | 147 (17.1%) |
Severe | 56 (30.4%) | 578 (46.9%) | 38 (28.8%) | 69 (39.4%) | 269 (31.3%) |
Young | | 586 (47.6%) | | 25 (14.3%) | 32 (3.7%) |
Middle aged | 33 (17.9%) | 493 (40.0%) | 26 (19.7%) | 129 (73.7%) | 243 (28.3%) |
Old | 154 (82.4%) | 172 (13.7%) | 110 (80.9%) | 23 (13.0%) | 600 (68.6%) |
Age (years) | 80 ± 10 | 44 ± 20 | 79 ± 11 | 55 ± 15 | 74 ± 15 |
Number of comorbidities | 4.5 ± 1.6 | 0.8 ± 0.8 | 5.5 ± 2.1 | 4.0 ± 1.2 | 2.5 ± 1.2 |
Each endotype had distinct rates of survival to discharge (Table 3). HE had the highest survival to discharge, followed by AA, HTN, RFH and HFA. Survival differences between all endotypes were statistically significant (HE 78.5%, AA 70.1%, HTN 51.4%, RFH 40.4%, HFA 27.8%, p < 0.05 for all comparisons). RFH had the lowest rate of neurosurgical intervention (13.2%) which was statistically lower than HFA (27.3%) and HTN (25.7%) which had the highest intervention rates (RFH vs HFA p = 0.033, RFH vs HTN p = 0.022). All other comparisons of intervention rates were non-significant between endotypes). For LOS, endotypes were split into long (HFA: 10.2 ± 7.2 days, RFH: 8.7 ± 6.0 days, AA: 10.6 ± 7.5 days,) and short (HE: 7.3 ± 6.9 days, HTN: 7.7 ± 6.4 days) LOS groups in where p < 0.05 for comparisons between endotypes across groups and p > 0.05 for comparisons of endotypes within a group.
Table 3
Outcomes of endotypes Outcomes of survival rate, neurosurgical intervention rate, and length of hospital stay (LOS) across the five stable comorbidity endotypes without any stratification. HFA: Heart failure and arrythmia, HE: Healthy, RFH: Renal failure with hypertension, AA: Alcohol abuse, HTN: Hypertension.
Outcome | HFA | HE | RFH | AA | HTN |
Alive | 52 | 984 | 55 | 124 | 450 |
Expired | 135 | 270 | 81 | 53 | 425 |
Survival rate (%) | 27.8 | 78.5 | 40.4 | 70.1 | 51.4 |
Neurosurgical Intervention | 51 | 269 | 18 | 41 | 225 |
No Neurosurgical Intervention | 136 | 985 | 118 | 136 | 650 |
Intervention rate (%) | 27.3 | 21.5 | 13.2 | 23.2 | 25.7 |
LOS (Days) | 10.2 ± 7.2 | 7.3 ± 6.9 | 8.7 ± 6.0 | 10.6 ± 7.5 | 7.7 ± 6.4 |
Survival to Discharge
Multivariable logistic regression models accounting for age, sex, and GCS showed that the endotypes HE, AA, and HTN had the highest survival rates (Table 4). Using the HE endotype as reference, AA (OR = 0.999, CI: 0.935–1.066) and HTN (OR = 0.962, CI: 0.919–1.007) had statistically similar survival rates compared to HE. HFA had the lowest survival rate (OR with respect to HE = 0.803, CI: 0.747–0.862) that was statistically different from all other endotypes. RFH had lower survival than HE (OR with respect to HE = 0.889, CI: 0.821–0.964) but higher than HFA (OR with respect to HFA = 1.108, CI: 1.011–1.214). Survival rates when accounting for age, sex, and GCS are as such: HE, AA, HTN > RFH > HFA. In contrast to the initial analyses, comparison of survival rates between HE, AA, and HTN become statistical insignificant when including age (OR per increase in age = 0.991, CI: 0.991–0.992) and GCS (OR per increase in GCS score = 1.028, CI: 1.024–1.031).
Table 4
Adjusted comparison of endotype survival outcomes Comparison of survival rates across the five stable comorbidity endotypes after adjustment by age, sex, and Glasgow coma scale (GCS) using multivariate logistic regression. HFA: Heart failure and arrythmia, HE: Healthy, RFH: Renal failure with hypertension, AA: Alcohol abuse, HTN: Hypertension. The base regression model uses the HE endotype as the reference for the endotype variable in survival odds ratio calculations. Models using the other endotypes as reference are also reported for comparison. Odds ratio for age, sex, and GCS omitted in these models as they remained constant for all models. Significance codes: "***": ≤ 0.001, "**": ≤ 0.01, "*": ≤ 0.05
Model | Variable | Odds Ratio | 95% CI | P-value |
HE as reference | HFA | 0.803 | 0.747–0.862 | < 0.001 *** |
RFH | 0.889 | 0.821–0.964 | 0.004 ** |
AA | 0.999 | 0.935–1.066 | 0.968 |
HTN | 0.962 | 0.919–1.007 | 0.095 |
Age (years) | 0.991 | 0.991–0.992 | < 0.001 *** |
GCS score | 1.028 | 1.024–1.031 | < 0.001 *** |
Male sex | 0.999 | 0.966–1.033 | 0.944 |
HFA as reference | HE | 1.246 | 1.159–1.339 | < 0.001 *** |
RFH | 1.108 | 1.011–1.214 | 0.0284 * |
AA | 1.244 | 1.139–1.359 | < 0.001 *** |
HTN | 1.198 | 1.123–1.279 | < 0.001 *** |
RFH as reference | HFA | 0.903 | 0.824–0.989 | 0.028 * |
HE | 1.125 | 1.038–1.219 | 0.004 ** |
AA | 1.123 | 1.021–1.235 | 0.017 * |
HTN | 1.082 | 1.004–1.166 | 0.040 * |
AA as reference | HFA | 0.804 | 0.736–0.878 | < 0.001 *** |
HE | 1.001 | 0.938–1.069 | 0.968 |
RFH | 0.890 | 0.810–0.979 | 0.017 * |
HTN | 0.963 | 0.899–1.032 | 0.288 |
HTN as reference | HFA | 0.834 | 0.782–0.891 | < 0.001 *** |
HE | 1.040 | 0.993–1.088 | 0.095 |
RFH | 0.924 | 0.857–0.996 | 0.040 * |
AA | 1.038 | 0.969–1.112 | 0.288 |
Additional investigation of these three endotypes based on stratification by age and GCS revealed that HE had higher survival in the middle-aged group compared to AA (80.3% vs 69.8%, p = 0.010) and was similar in all age groups in comparison to HTN (Table 5). However, when further sub-stratifying age groups by GCS, some substrata comparisons become significant: HE has a higher survival rate than AA for the mild GCS group in both young (97.5% vs 83.3%, p = 0.043) and middle-aged groups (94.1% vs 70.0%, p < 0.001). Although survival was not different across age strata between HE and HTN, sub-stratification by GCS revealed that HE had better survival for the mild (94.1% vs 87.2%, p = 0.043) and severe (66.7% vs 52.6%, p = 0.043) GCS groups within the middle-aged group.
Table 5
Stratified endotype survival outcomes Comparison of survival after stratification by both injury severity and for Healthy (HE), Alcohol abuse (AA), and Hypertension (HTN) endotypes. Only HE, AA, and HTN were chosen for additional stratified analysis as results from regression with adjustment by age and Glasgow coma scale (GCS) differed from initial results and due to significant variations of GCS and age distributions between these three endotypes. Young: 16–39, Middle-aged: 40–69, Old: 70+. Mild: GCS 14–15, Moderate: GCS 9–13, Severe: GCS 3–8.
Age | TBI Severity | Outcome | HE | AA | HTN |
Young | Mild | Alive | 193 | 5 | 6 |
Expired | 5 | 1 | 1 |
Survival rate (%) | 97.5 | 83.3 | 85.7 |
Moderate | Alive | 84 | 4 | 2 |
Expired | 5 | 1 | 0 |
Survival rate (%) | 94.4 | 80.0 | 100.0 |
Severe | Alive | 237 | 11 | 18 |
Expired | 56 | 3 | 4 |
Survival rate (%) | 80.9 | 78.6 | 81.8 |
Middle-aged | Mild | Alive | 177 | 35 | 109 |
Expired | 11 | 15 | 16 |
Survival rate (%) | 94.1 | 70.0 | 87.2 |
Moderate | Alive | 73 | 21 | 28 |
Expired | 12 | 7 | 7 |
Survival rate (%) | 85.9 | 75.0 | 80.0 |
Severe | Alive | 140 | 33 | 40 |
Expired | 70 | 16 | 36 |
Survival rate (%) | 66.7 | 67.3 | 52.6 |
Old | Mild | Alive | 38 | 10 | 158 |
Expired | 30 | 5 | 153 |
Survival rate (%) | 55.9 | 66.7 | 50.8 |
Moderate | Alive | 11 | 1 | 39 |
Expired | 15 | 1 | 71 |
Survival rate (%) | 42.3 | 50.0 | 35.5 |
Severe | Alive | 19 | 3 | 41 |
Expired | 56 | 3 | 130 |
Survival rate (%) | 25.3 | 50.0 | 24.0 |
Neurosurgical Intervention
In comparison to HE, HFA (OR = 1.146, CI: 1.067–1.231) and HTN (OR = 1.123, CI: 1.074–1.175) had the highest rate of neurosurgical intervention while RFH (OR = 1.005, CI: 0.928–1.089) and AA (OR = 1.037, CI: 0.972–1.107) had statistically similar rates of intervention after adjusting for age, sex, and GCS (Table 6). Overall, results from the multivariable logistic regression splits the endotypes into the low (HE, RFH, AA) and high (HFA, HTN) intervention groups where rates are statistically different compared to endotypes of the other group for comparisons) while being statistically similar to endotypes of the same group.
Table 6
Adjusted comparison of endotype intervention rates Comparison of rates of neurosurgical intervention across the five stable comorbidity endotypes after adjustment by age, sex, and Glasgow coma scale (GCS) using multivariate logistic regression. The base regression model uses the Healthy (HE) endotype as the reference for the endotype variable in intervention rate odds ratio calculations. Models using the other endotypes as reference are also reported for comparison. Odds ratio for age, sex, and GCS omitted in these models as they remained constant for all models. HFA: Heart failure and arrythmia, RFH: Renal failure with hypertension, AA: Alcohol abuse, HTN: Hypertension. Significance codes: "***": ≤ 0.001, "**": ≤ 0.01, "*": ≤ 0.05
Model | Variable | Odds Ratio | 95% CI | P-value |
HE as reference | HFA | 1.146 | 1.067–1.231 | < 0.001 *** |
RFH | 1.005 | 0.928–1.089 | 0.893 |
AA | 1.037 | 0.972–1.107 | 0.272 |
HTN | 1.123 | 1.074–1.175 | < 0.001 *** |
Age (years) | 0.999 | 0.998–1.000 | 0.034 * |
GCS score | 0.978 | 0.975–0.982 | < 0.001 *** |
Male sex | 1.063 | 1.029–1.099 | < 0.001 *** |
HFA as reference | HE | 0.872 | 0.812–0.937 | < 0.001 *** |
RFH | 0.877 | 0.801–0.961 | 0.005 ** |
AA | 0.905 | 0.829–0.987 | 0.025 * |
HTN | 0.980 | 0.918–1.046 | 0.543 |
RFH as reference | HFA | 1.140 | 1.041–1.248 | 0.005 ** |
HE | 0.995 | 0.918–1.077 | 0.893 |
AA | 1.032 | 0.939–1.134 | 0.519 |
HTN | 1.117 | 1.037–1.204 | 0.004 ** |
AA as reference | HFA | 1.105 | 1.013–1.206 | 0.025 * |
HE | 0.964 | 0.903–1.029 | 0.272 |
RFH | 0.969 | 0.882–1.065 | 0.519 |
HTN | 1.083 | 1.011–1.16 | 0.022 * |
HTN as reference | HFA | 1.020 | 0.956–1.089 | 0.543 |
HE | 0.890 | 0.851–0.931 | < 0.001 *** |
RFH | 0.895 | 0.831–0.964 | 0.004 ** |
AA | 0.923 | 0.862–0.989 | 0.022 * |
Length of Stay
Using ANCOVA to adjust for the covariates of age, sex, and GCS, HE had shorter LOS than HFA (difference = 4.6 ± 0.9 days, p < 0.001), RFH (difference = 3.3 ± 1.1 days, p = 0.015), and AA (different = 4.8 ± 0.9 days, p < 0.001) but had similar LOS compared to HTN (difference = 1.2 ± 0.6 days, p = 0.214) (Table 7). HFA and AA had almost identical LOS (difference = 0.2 ± 1.2 days, p = 1.000). LOS for the RFH endotype was statistically comparable to every other endotype (p > 0.05 for all comparisons) except HE. Overall, HE and HTN had the shortest LOS while HFA and AA had the longest LOS of the endotypes after accounting for age, sex, and GCS.
Table 7
Adjusted comparisons for endotype length of stay Comparison length of stay (LOS) across the five stable comorbidity endotypes after adjustment by age, sex, and Glasgow coma scale (GCS) using analysis of covariance (ANCOVA). ANCOVA model output and results of Tukey post-hoc comparisons for LOS between each pair combination of endotypes are reported below. HFA: Heart failure and arrythmia, HE: Healthy, RFH: Renal failure with hypertension, AA: Alcohol abuse, HTN: Hypertension. Significance codes: "***": ≤ 0.001, "**": ≤ 0.01, "*": ≤ 0.05
Variable | Df | Sum of Squares | Mean Square | F-value | P-value |
Endotype | 4 | 4430 | 1108 | 9.896 | < 0.001 *** |
Age (years) | 1 | 1881 | 1881 | 16.807 | < 0.001 *** |
GCS | 1 | 13656 | 13656 | 122.012 | < 0.001 *** |
Sex | 1 | 593 | 593 | 5.296 | 0.022 * |
Residuals | 2574 | 288080 | 112 | | |
Comparison Hypothesis | LOS difference (days) | Standard Error (days) | t-value | P-value |
HFA - HE = 0 | 4.6 | 0.9 | 4.815 | < 0.001 *** |
RFH - HE = 0 | 3.3 | 1.1 | 3.101 | 0.015 * |
AA - HE = 0 | 4.8 | 0.9 | 5.556 | < 0.001 *** |
HTN - HE = 0 | 1.2 | 0.6 | 2.083 | 0.214 |
RFH - HFA = 0 | -1.3 | 1.2 | -1.059 | 0.816 |
AA - HFA = 0 | 0.2 | 1.2 | 0.210 | 1.000 |
HTN - HFA = 0 | -3.3 | 0.9 | -3.845 | 0.001 ** |
AA - RFH = 0 | 1.5 | 1.3 | 1.214 | 0.729 |
HTN - RFH = 0 | -2.0 | 1.0 | -2.055 | 0.226 |
HTN - AA = 0 | -3.6 | 0.9 | -3.912 | < 0.001 *** |
Data Availability: |
All code used for analysis and visualization of our data is available on GitHub: https://github.com/SteveHQiu/TBIClustering |
Data used in our analysis were extracted from the Medical Information Mart for Intensive Care III (MIMIC-III) dataset which is freely-available database: https://physionet.org/content/mimiciii/1.4/ |
Predictive Value of Comorbidities and Endotypes for Survival
Logistic regression using all 30 Elixhauser comorbidity categories to predict survival to discharge accounts for ~ 15.3% of the variation in outcome (F(30, 2598) = 15.59, R2 = 0.153, p < 0.001) and had an AUC of 0.73. A simpler logistic regression using the five derived comorbidity endotypes accounted for 11.9% of the variation in survival (F(4, 2624) = 88.61, R2 = 0.119, p < 0.001) with an AUC of 0.69. Combining the base 30 Elixhauser categories with the endotypes in a full logistic regression model explains 17.0% of the variation in survival (F(34, 2594) = 15.57, R2 = 0.170, p < 0.001) and has an AUC of 0.74. The decrease in model error for predicting survival to discharge seen between the base logistic regression (including only the comorbidities) and the full model (including both comorbidities and endotypes) is statistically significant (F(4, 2594) = 13.23, p < 0.001).