A Novel Risk Score to Predict Post-Trauma Mortality in Nonagenarians.

BACKGROUND: Nonagenarians represent a rapidly growing age group who often have functional limitations and multiple comorbidities, predisposing them to trauma. AIMS: The purpose of this study was to identify patient characteristics, hospital complications, and comorbidities that predict in-hospital mortality in the nonagenarian population following trauma. We also sought to create a scoring system using these variables. SETTINGS AND
DESIGN: This study was a retrospective chart review.
METHODS: We reviewed the medical records of 548 nonagenarian trauma patients admitted to two Level I trauma centers from 2006 to 2015. Statistical analysis was performed using logistic regression and a machine learning model, which calculated significant variables and computed a scoring system.
RESULTS: The in-hospital mortality rate was 7.1% (n = 39). Significant predictors of mortality were cardiac comorbidity, neuro-concussion, New Injury Severity Score (ISS) 16+, striking an object, ISS 25-75, and pulmonary and cardiac complications. Significant variables were assigned a numeric value. A score of 5+ carried a 41.1% mortality risk, 79% sensitivity, and 91% specificity. A score of 10+ had an associated 81.8% mortality risk with 31% specificity and 99% sensitivity.
CONCLUSIONS: Our findings identified reliable predictors of mortality in nonagenarian population posttrauma. The scoring system performs with good specificity and sensitivity and incrementally correlates with mortality risk.

INTRODUCTION

More than 88 million Americans are estimated to be 65 years of age and older by 2035, comprising a quarter of the population and for the first time, outnumbering those under the age of 18.[1] With continued population aging, a significant number of older adults will be nonagenarians (>90 years old). Patients in this population age group tend to have visual and hearing impairments, multiple comorbidities, cognitive decline, and higher rates of frailty and polypharmacy,[23] all of which contribute to the occurrence of traumatic events such as falls.

Nonagenarian falls is the most common mechanism of injury to adult trauma centers.[4] When combined with motor vehicle accidents and other causes of trauma, these patients are partially responsible for approximately 10% of the presenting cases and have the highest case fatality rate.[4]

An accurate prediction of which patients are susceptible to a higher risk of mortality can help guide clinicians' decision-making and prompt earlier discussion of goals of care. Previous scoring systems such as the Physiologic and Operative Severity Score for Enumeration of Mortality and Morbidity (POSSUM) and Portsmouth POSSUM were found to be unreliable and overpredicted mortality, especially in high-risk groups.[5] Several factors including presenting systolic blood pressure, Glasgow Coma Scale (GCS), admission pH and lactate, Geriatric Trauma Outcome Score, and the need for mechanical ventilation have been shown to predict mortality, but in a younger geriatric population,[67] and how these factors relate to the nonagenarian patient remains largely unknown.

The primary objectives of this study were to identify the predictors of in-hospital posttrauma mortality among nonagenarians and establish a risk assessment scoring system.

METHODS

A retrospective chart review was conducted on 538 nonagenarian trauma patients admitted to two Level I New York City trauma centers from 2006 to 2015. All patients admitted under the trauma service, and who were over the age of 90, were included in the study. Patients admitted to another service were excluded from the study. Institutional Review Board approval and informed consent were waived as this study was a retrospective chart review and posed little risk to patients.

Patients were assigned an Injury Severity Score (ISS) which is a sum of squares of the highest Abbreviated Injury Scales grade in the three most severely injured body regions with possible values from 1 to 75. An ISS score can predict mortality and best regarded as a categorical variable, but there is variability among studies regarding the number of categories and severity classification.[8] A modified ISS (New ISS [NISS]), based on the three most severe injuries, has been shown in some studies to have better predictive value than the ISS and was assigned to each patient.[9] Other independent variables included admitting diagnosis, comorbidities, complications, vitals, basic demographics, and mechanism of injury. The ten most common comorbidities were hypertension, coronary artery disease (CAD), arrhythmia, congestive heart failure, diabetes, chronic obstructive pulmonary disease, asthma, renal related, psychiatric related, cerebrovascular accident (CVA), and anemia. Complications included infection (urinary tract infection, sepsis from any source), pulmonary (acute respiratory distress syndrome, respiratory distress, pneumonia, pleural effusions), hematological (anemia, coagulation disorders, gastrointestinal bleed), cardiac (myocardial infarction [MI], CAD, arrhythmia), renal diagnosis, deep-vein thrombosis/pulmonary embolism, and decubitus ulcer, small bowel obstruction, and seizures. The outcome of interest was in-hospital trauma service mortality.

We described univariate associations between each of the hypothesized predictors and the outcome and mortality, using χ2 tests for categorical variables and Kruskal–Wallis tests for continuous variables, with statistical significance considered at P <0.05. Model development was based on the tools for prediction developed under the paradigm of machine learning, which uses concepts from both computer science and statistics. As the goal was to develop a simple additive model with a limited number of parameters for use in clinical practice, model fitting involved the use of logistic regression with the Least Absolute Shrinkage and Selection Operator (LASSO). This technique is a form of penalized regression that introduces a penalty to the model deviance, forcing the sum of the absolute value of the coefficients to be less than a constant.[10]

The effect of this penalty is to shrink several coefficients to zero while optimizing a model selection criterion such as area under the receiver operating characteristic (ROC) curve (AUC) [Figure 1].

Figure 1

Receiver operating characteristic curves

We used 10-fold cross-validation on the training set to select the parameter that maximizes AUC. This process was performed for LASSO logistic regression models with a range of values, and the value that produced the largest cross-validation AUC was selected and fit to the entire training set. We then computed sensitivity, specificity, negative predictive value (NPV), positive predictive value (PPV), and Cohen's as measures of model performance when the final model was applied to the test set.

To compute a simplified risk score, we used bootstrap resampling on the entire dataset with 1000 replications and converted the LASSO coefficients to integers by multiplying the coefficients of a range of scalar values, and then rounding to the nearest integer, with this process repeated for each bootstrap replication. We selected the scalar parameter that maximized the median of the bootstrap AUC distribution and calculated the final risk score using that parameter on the entire dataset. We again computed sensitivity, specificity, NPV, PPV, and Cohen's to evaluate the predictive ability of the risk score when applied to the test set. These parameters were compared to those obtained in the full model.

RESULTS

Sample demographics, complications, comorbidities, fall mechanisms, and injury severity scores are presented in Tables 1-3 along with their corresponding univariate associations with mortality. The sample had a mean age of 93.8 ± 3.0 years of age with varying fall mechanisms, of which 46 (8.3%) suffered a fall from height, 317 (56.8%) fell from standing, 33 (5.9%) struck a sharp object, and 162 (29%) had an unspecified fall.

Table 1

Sample demographics

	Overall	Alive	Died	P
n	538	499	39
Age - mean±SD	93.8±3.0	93.7±3.0	94.2±3.3	0.350
Number of comorbidities - mean±SD	1.5±0.9	1.5±0.9	1.6±0.8	0.298
Number of complications - mean±SD	1.1±0.3	1.1±0.3	1.2±0.4	0.00269
Glascow Coma Scale - mean±SD	14.6±1.2	14.8±0.8	13.1±2.7	<0.001
Pulse - mean±SD	81.9±16.4	81.2±15.1	91.3±26.7	<0.001
Systolic blood pressure - mean±SD	153.0±29.5	153.4±29.1	148.3±34.2	0.350
Sex (%)
Female	411 (76.4)	389 (78.0)	22 (56.2)	0.010
Male	127 (23.6)	110 (22.0)	17 (43.8)
Race/ethnicity (%)
Asian	41 (7.7)	36 (7.4)	5 (12.9)	0.532
Black	9 (1.6)	9 (1.7)	0 (0.0)
Other	34 (6.4)	33 (6.6)	1 (3.2)
White	460 (84.3)	430 (84.3)	38 (83.9)

SD: Standard deviation

Table 3

Fall mechanism and Injury Severity Scores

	Overall	Alive	Died	P
Fall mechanism (%)
Blunt fall from height	46 (8.3)	41 (7.9)	5 (13.8)	0.175
Fall from standing	317 (56.8)	302 (58.2)	15 (37.9)
Striking sharp object	33 (5.9)	29 (5.6)	4 (10.3)
Unspecified fall	162 (29.0)	148 (28.3)	14 (37.9)
ISS score (%)
1-3	86 (15.4)	84 (16.1)	2 (6.2)	<0.001
4-8	210 (37.6)	205 (39.6)	5 (12.5)
9-15	208 (37.2)	189 (36.4)	9 (46.9)
16-24	42 (7.5)	37 (7.1)	5 (12.5)
25-75	13 (2.3)	4 (0.7)	9 (21.9)
NISS 16 or over (%)	82 (14.7)	63 (12.2)	19 (46.9)	<0.001

ISS: Injury Severity Score, NISS: New ISS

Variables indicating statistically significant relationships with mortality in univariate analysis included sex (P = 0.010), pulmonary complication (P < 0.001), cardiac complication (P < 0.001), neurosurgical diagnosis (P < 0.001), neuro-head diagnosis (P = 0.003), neuro-concussion (P < 0.001), number of complications (P = 0.003), ISS score (P < 0.001), NISS score 16 or over (P < 0.001), GCS (P < 0.001), and pulse (P = 0.001).

Figure 2 plots the relationship between lambda and AUC. The value of lambda that maximized the AUC in cross-validation was 0.007, selected from the range of 0, 0.102, the latter being the maximum shrinkage parameter. Results of the model with lambda = 0.007 fit to the full training dataset are presented in Figure 3. This model contained coefficients for pulmonary complication (β = 2.0), cardiac complication (β = 1.7), cardiac comorbidity (β = 0.2), neuro-concussion (β = 0.1), ISS 25–75 (β = 1.3), NISS 16+ (β = 0.4), and striking sharp object (β = 0.6). The full model had an AUC (95% confidence interval [CI]) of 0.89 (0.8–0.99) when applied to the training dataset and 0.93 (0.88–0.98) when applied to the test dataset.

Figure 2

Relationship between lambda and area under the receiver operating characteristic curve (model selection criterion)

Figure 3

Full model coefficients

The resulting risk score is composed of the following: add 1 point each for any cardiac comorbidity or any neuro-concussion, 2 points each for NISS 16+ or striking a sharp object, 4 points for ISS 25–75, and 5 points each for any pulmonary or cardiac complication [Table 4]. Figure 4 compares the performance of the full model to the simplified risk score, each applied to the test set, using ROC curves. The risk score can retain a nearly equivalent discriminatory power AUC (95% CI: 0.91 [0.86–0.97]), compared to the full model (0.93 [0.88–0.98]). Using a score cutoff of 6+, the score attains a sensitivity of 0.76, specificity of 0.93, PPV of 0.47, NPV of 0.98, and Cohen's κ (95% CI: 0.54 [0.40–0.68]). Evaluation metrics for each additional possible cutoff of the risk score are presented in Table 5.

Table 4

Scoring system

Points	Variable
1	Cardiac comorbidity
1	Neuro-concussion
2	NISS: 16+
2	Striking sharp object
4	ISS: 25-75
5	Pulmonary complication
5	Cardiac complication

ISS: Injury Severity Score, NISS: New ISS

Figure 4

Area under the receiver operating characteristic curve for risk scores based on a range of rounding parameters, using 1000 bootstrap resamples

Table 5

Risk score evaluation metrics

Score	Risk (%)	Sensitivity	Specificity	PPV	NPV	κ
1+	8.5	1.00	0.18	0.09	1.00	0.03 (0.02-0.04)
2+	17.9	0.93	0.67	0.18	0.99	0.21 (0.13-0.28)
3+	25.3	0.86	0.80	0.25	0.99	0.32 (0.21-0.42)
4+	32.9	0.83	0.87	0.33	0.99	0.41 (0.29-0.53)
5+	41.1	0.79	0.91	0.41	0.98	0.49 (0.36-0.63)
6+	46.8	0.76	0.93	0.47	0.98	0.54 (0.40-0.68)
7+	50.0	0.52	0.96	0.50	0.96	0.47 (0.31-0.63)
8+	60.9	0.48	0.98	0.61	0.96	0.51 (0.34-0.68)
9+	70.6	0.41	0.99	0.71	0.96	0.50 (0.31-0.68)
10+	81.8	0.31	0.99	0.82	0.95	0.43 (0.23-0.62)
11+	88.9	0.28	1.00	0.89	0.95	0.40 (0.20-0.60)
12+	87.5	0.24	1.00	0.88	0.95	0.36 (0.16-0.56)
13+	100.0	0.17	1.00	1.00	0.94	0.28 (0.09-0.47)

NPV: Negative predictive value, PPV: Positive predictive value

DISCUSSION

The findings of this study resulted in an easily computed in-hospital mortality risk score for nonagenarian trauma patients.

Although nonagenarians require more comprehensive care in the postoperative period due to their multiple chronic illnesses and susceptibility to postoperative complications, functional decline, and mortality, few studies have attempted to predict their mortality risk posttrauma. A strength of this study was the large, diverse multicentric sample of patients increasing the validity. Advanced age may inherently predispose nonagenarians to a poor prognosis after a traumatic event; nevertheless, identifying patients most at risk for mortality can expedite the provision of better quality of care including resources and arranging advanced directives. This study identified pulmonary and cardiac complications as postoperative complications that predisposed to mortality, and nonagenarian patients should receive care according to the best practice guidelines[11] to minimize or prevent these complications.

Aging has a profound impact on the physiologic capacity to respond to injury, progressively impairing adaptive and homeostatic mechanisms. As a result, variables integral to the prognosis of each age group in the geriatric trauma population require unique consideration and exploration.[12] Mortality in the centenarian and nonagenarian population posttrauma has been shown to be associated with age in addition to mean ISS score and number of comorbidities.[13] Our findings regarding the nonagenarian population exclusively confirm these predictors of mortality. Other risk factors such as severity of injury scores and pneumonia had predictive value in the setting of trauma, whereas systolic blood pressure (BP) had no predictive value as reported in previous studies investigating nonagenarian mortality.[14]

Sex, GCS, and pulse were found to be independently associated with mortality after univariate analysis. For patients older than 70 years of age, which includes our population of interest, increased mortality risk was found to correspond with a systolic BP lower than 110. For younger patients, a systolic BP <90 increases the probability of mortality in trauma.[15] Our study showed that a mean pulse of 91.3 and not systolic BP were associated with increased mortality, whereas a mean pulse of 81.5 showed risk reduction. Our findings were unanticipated when considering the Cushing reflex, a physiologic triad of bradycardia, hypertension, and respiratory irregularity, representing increased intracranial pressure and elevated mortality risk.[16] However, bradycardia has been shown to be a weak prognostic indicator,[17] and a presenting tachycardia is still associated with significant mortality risk.[18] A heart rate outside 70 and 89 beats per minute has been found to be associated with increased mortality,[19] in congruence with our study findings. The Cushing reflex, hemodynamic responses, and their effects on mortality require further investigation in the nonagenarian trauma patients.

A lower GCS was associated with an increased mortality. Nonagenarians were more likely to die with a GCS of 13.1 and a standard deviation of 2.7, than with a GCS of 14.8 and a standard deviation of 0.8. This small but significant difference was unexpected as other studies show a greater range in GCS before there is an increase in mortality. One meta-analysis found increased mortality at a GCS of 9–12, rather than 13 in an older adult population. These studies included a wide range of “elderly” patients, while ours focused on nonagenarians.[1520] Furthermore, older adults tend to present with higher GCS scores despite suffering more severe injuries[21] when compared to younger patients, coinciding with our results which showed significant mortality risk even with a relatively high GCS. These findings assist in the triage of older adult trauma patients and signify the unreliability of relying solely on a low GCS to determine injury severity. Instead of using a pooled score, future studies may investigate specific deficits and their impact on mortality.

The only comorbidity that calculates into the risk score is cardiac, which includes any arrhythmias, MIs, or CAD. Grossman et al. identified hepatic disease, renal disease, and cancer as preexisting conditions that significantly impact mortality in the elderly trauma patient. Grossman et al. examined all trauma patients >65 years of age.[22] Patients who have a diagnosis of cancer and live to be at least 65 years old were found to die from their injuries more often and would have increased complications from that cancer, such as pathologic fractures, or no escalation of care due to the patient's diagnosis. Patients who have survived to 90 years of age may have lower rates and less severe cancers and may suffer mortality due to even more chronic conditions, i.e., cardiac arrhythmia in the setting of trauma. Most patients with advanced cirrhosis, portal hypertension, or other parenchymal disease do not generally live to be 90 years old.[222324]

It has been shown in previous studies that anticoagulation alone (specifically warfarin) increases mortality in elderly trauma patients. We did not look at the effects of anticoagulation on nonagenarians with CAD. Although it is likely that many nonagenarians with CAD would also be on anticoagulation, it is unknown to the authors if the specific comorbidity of CAD alone is responsible for increasing mortality, or if it is the association of CAD with anticoagulation.[2225] This is an area where further study can be performed.

Pulmonary and cardiac complications are not independently studied in the elderly literature, but as a constellation of vital signs and triage findings. These complications significantly improved the predictive value of our score model, more so than any other variable, confirming the profound impact they have in predicting the mortality of the nonagenarian trauma patient.[1522262728]

Several limitations are inherent to our study findings. There is selection bias due to the retrospective design, but given the large sample size, our results have sufficient external validity and can be extrapolated to the larger population. Data available to the research team was limited and could not be expanded to include the cause of death, medications, length of stay in hospital, and outcomes postdischarge. Furthermore, if patients were transferred to medical services other than the trauma team, they were excluded from the sample. These variables would have provided a more accurate clinical representation of the postoperative nonagenarian patients. Despite these omissions, our model has good predictive value and can discriminate between survivors and nonsurvivors. This investigation aims to assess a specific outcome among nonagenarian trauma patients and serves to expand the knowledge base of a novel research area.

The “Mortality Index” is easy and quick to calculate. It can be utilized on admission to intensive care units, surgical wards, or even in the trauma bay by assessing for cardiac comorbidities, calculation of an ISS/NISS score, admitting diagnosis, and any complications that occurred. The appropriate values can then be added and correlated with the percent mortality.

CONCLUSIONS

This study found that comorbidities, increased ISS/NISS score, and cardiac/pulmonary complications increase mortality in the nonagenarian trauma population. In some cases, small quantitative changes in these variables led to significantly worse outcomes with regard to mortality and may be due to intricate physiologic responses to trauma.

The “Mortality Index” stratifies these significant parameters into a model that predicts mortality in the nonagenarian trauma patient. We feel that this tool is an effective adjunct to the trauma surgeon's clinical skills and experience. It can be used for end-of-life discussions, rationally predicting the treatment outcomes and putting the true percentage of mortality in the hands of the treating physician.

Financial support and sponsorship

Nil.

Conflicts of interest

There are no conflicts of interest.

Table 2

Sample complications and comorbidities

	Overall	Alive	Died	P
Complications (%)
Infection	48 (8.9)	40 (8.1)	8 (18.8)	0.0886
Pulmonary complication	54 (10.0)	31 (6.1)	23 (59.4)	<0.001
Cardiac complication	17 (3.2)	9 (1.7)	8 (21.9)	<0.001
Comorbidities (%)
Cardiac comorbidity	364 (67.8)	332 (66.7)	32 (81.2)	0.135
Psych/dementia	97 (17.9)	88 (17.6)	9 (21.9)	0.713
Orthopedic procedure	207 (37.2)	198 (38.4)	9 (21.9)	0.0947
Proximal orthopedic procedure	133 (23.8)	125 (24.2)	8 (18.8)	0.630
Distal orthopedic procedure	77 (13.8)	76 (14.7)	1 (3.1)	0.120
Proximal orthopedic diagnosis	184 (33.0)	176 (34.1)	8 (18.8)	0.114
Distal orthopedic diagnosis	90 (16.1)	89 (17.1)	1 (3.1)	0.0682
Neurosurgical diagnosis	229 (40.9)	199 (38.2)	30 (75.0)	<0.001
Neuro-head diagnosis	61 (10.9)	50 (9.5)	11 (28.1)	0.00311
Neuro-spinal diagnosis	67 (11.8)	59 (11.3)	8 (18.8)	0.329
Neuro-concussion	180 (32.3)	155 (29.9)	25 (62.5)	<0.001