Outcome Prediction in Patients with Severe Traumatic Brain Injury Using Deep Learning from Head CT Scans

Background After severe traumatic brain injury (sTBI), physicians use long-term prognostication to guide acute clinical care yet struggle to predict outcomes in comatose patients. Purpose To develop and evaluate a prognostic model combining deep learning of head CT scans and clinical information to predict long-term outcomes after sTBI. Materials and Methods This was a retrospective analysis of two prospectively collected databases. The model-building set included 537 patients (mean age, 40 years ± 17 [SD]; 422 men) from one institution from November 2002 to December 2018. Transfer learning and curriculum learning were applied to a convolutional neural network using admission head CT to predict mortality and unfavorable outcomes (Glasgow Outcomes Scale scores 1-3) at 6 months. This was combined with clinical input for a holistic fusion model. The models were evaluated using an independent internal test set and an external cohort of 220 patients with sTBI (mean age, 39 years ± 17; 166 men) from 18 institutions in the Transforming Research and Clinical Knowledge in Traumatic Brain Injury (TRACK-TBI) study from February 2014 to April 2018. The models were compared with the International Mission on Prognosis and Analysis of Clinical Trials in TBI (IMPACT) model and the predictions of three neurosurgeons. Area under the receiver operating characteristic curve (AUC) was used as the main model performance metric. Results The fusion model had higher AUCs than did the IMPACT model in the prediction of mortality (AUC, 0.92 [95% CI: 0.86, 0.97] vs 0.80 [95% CI: 0.71, 0.88]; P < .001) and unfavorable outcomes (AUC, 0.88 [95% CI: 0.82, 0.94] vs 0.82 [95% CI: 0.75, 0.90]; P = .04) on the internal data set. For external TRACK-TBI testing, there was no evidence of a significant difference in the performance of any models compared with the IMPACT model (AUC, 0.83; 95% CI: 0.77, 0.90) in the prediction of mortality. The Imaging model (AUC, 0.73; 95% CI: 0.66-0.81; P = .02) and the fusion model (AUC, 0.68; 95% CI: 0.60, 0.76; P = .02) underperformed as compared with the IMPACT model (AUC, 0.83; 95% CI: 0.77, 0.89) in the prediction of unfavorable outcomes. The fusion model outperformed the predictions of the neurosurgeons. Conclusion A deep learning model of head CT and clinical information can be used to predict 6-month outcomes after severe traumatic brain injury. © RSNA, 2022 Online supplemental material is available for this article. See also the editorial by Haller in this issue.

Risk Factors and Incidence of Epilepsy after Severe Traumatic Brain Injury

We determined the incidence of post-traumatic epilepsy after severe traumatic brain injury. Of 392 patients surviving to discharge, cumulative incidence of post-traumatic epilepsy was 25% at 5 years and 32% at 15 years, an increase compared with historical reports. Among patients with one late seizure (>7 days post-trauma), the risk of seizure recurrence was 62% after 1 year and 82% at 10 years. Competing hazards regression identified age, decompressive hemicraniectomy, and intracranial infection as independent predictors of post-traumatic epilepsy. Patients with severe traumatic brain injury and a single late post-traumatic seizure will likely require long-term antiseizure medicines. ANN NEUROL 2022;92:663-669.

Performance of CRASH and IMPACT Prognostic Models for Traumatic Brain Injury at 12 and 24 Months Post-Injury

The Corticoid Randomization after Significant Head Injury (CRASH) and International Mission for Prognosis and Analysis of Clinical Trials (IMPACT) prognostic models are the most reported prognostic models for traumatic brain injury (TBI) in the scientific literature. However, these models were developed and validated to predict 6-month unfavorable outcome and mortality, and growing evidence supports continuous improvements in functional outcome after severe TBI up to 2 years post-injury. The purpose of this study was to evaluate CRASH and IMPACT model performance beyond 6 months post-injury to include 12 and 24 months post-injury. Discriminative validity remained consistent over time and comparable to earlier recovery time points (area under the curve = 0.77-0.83). Both models had poor fit for unfavorable outcomes, explaining less than one quarter of the variation in outcomes for severe TBI patients. The CRASH model had significant values for the Hosmer-Lemeshow test at 12 and 24 months, indicating poor model fit past the previous validation point. There is concern in the scientific literature that TBI prognostic models are being used by neurotrauma clinicians to support clinical decision making despite the goal of the models' development being to support research study design. The results of this study indicate that the CRASH and IMPACT models should not be used in routine clinical practice because of poor model fit that worsens over time and the large, unexplained variance in outcomes.

Complications from Multimodal Monitoring Do not Affect Long-Term Outcomes in Severe Traumatic Brain Injury

BACKGROUND

Invasive neuromonitoring is a mainstay of modern management of severe traumatic brain injury (TBI). Complication rates of neuromonitor placement are widely reported, but their effects on long-term outcomes are less studied. We evaluated the association of neuromonitor complications on long-term outcomes in a prospective severe TBI cohort.

METHODS

We reviewed 599 patients with severe TBI from November 2002 through 2018 for neuromonitor-associated hemorrhage and infection. We compared outcome differences between patients with and without neuromonitoring-associated complications using the Glasgow Outcomes Scale (GOS) at 3, 6, 12, and 24 months post trauma. When analyzing neuromonitoring infections, we removed all patients who expired before discharge as early mortality was associated with reduced infection rates.

RESULTS

Neuromonitor-associated hemorrhage occurred in 62 out of 534 patients with post placement imaging (11.6%) and was increased in patinets who underwent a craniotomy (24% vs. 11%, P = 0.005). Clinical outcomes did not differ in patients with neuromonitor-associated hemorrhage. Neuromonitor-associated infection occurred in 30 of 389 patients (7.7%) who survived to discharge. Infection was associated with worse outcomes at 3 months (P = 0.03), where the proportion of patients with favorable outcomes (P = 0.02) was decreased despite similar mortality (P = 0.24). Patients with an infection recovered by 6 months, at which point there were no differences in total GOS or rates of favorable outcomes then or at later time points (P > 0.26). Neuromonitor-associated infection was associated with increased length of stay (P = 0.01) and depressed skull fractures (P = 0.03) but did not affect rates of shunting (P = 0.99).

CONCLUSIONS

Complications of neuromonitoring in severe TBI are associated with delayed recovery but not long-term outcomes.

Prognostic Models for Traumatic Brain Injury Have Good Discrimination but Poor Overall Model Performance for Predicting Mortality and Unfavorable Outcomes

BACKGROUND

The most extensively validated prognostic models for traumatic brain injury (TBI) are the Corticoid Randomization after Significant Head Injury (CRASH) and International Mission on Prognosis and Analysis of Clinical Trials (IMPACT). Model characteristics outside of area under the curve (AUC) are rarely reported.

OBJECTIVE

To report the discriminative validity and overall model performance of the CRASH and IMPACT models for prognosticating death at 14 days (CRASH) and 6 months (IMPACT) and unfavorable outcomes at 6 months after TBI.

METHODS

This retrospective cohort study included prospectively collected patients with severe TBI treated at a single level I trauma center (n = 467). CRASH and IMPACT percent risk values for the given outcome were computed. Unfavorable outcome was defined as a Glasgow Outcome Scale-Extended score of 1 to 4 at 6 months. Binary logistic regressions and receiver operating characteristic analyses were used to differentiate patients from the CRASH and IMPACT prognostic models.

RESULTS

All models had low R 2 values (0.17-0.23) with AUC values from 0.77 to 0.81 and overall accuracies ranging from 72.4% to 78.3%. Sensitivity (35.3-50.0) and positive predictive values (66.7-69.2) were poor in the CRASH models, while specificity (52.3-53.1) and negative predictive values (58.1-63.6) were poor in IMPACT models. All models had unacceptable false positive rates (20.8%-33.3%).

CONCLUSION

Our results were consistent with previous literature regarding discriminative validity (AUC = 0.77-0.81). However, accuracy and false positive rates of both the CRASH and IMPACT models were poor.

Hydrocephalus and Cerebrospinal Fluid Analysis Following Severe Traumatic Brain Injury: Evaluation of a Prospective Cohort

The development of hydrocephalus after severe traumatic brain injury (TBI) is an under-recognized healthcare phenomenon and can increase morbidity. The current study aims to characterize post-traumatic hydrocephalus (PTH) in a large cohort. Patients were prospectively enrolled age 16-80 years old with Glasgow Coma Scale (GCS) score ≤8. Demographics, GCS, Injury Severity Score (ISS), surgery, and cerebrospinal fluid (CSF) were analyzed. Outcomes were shunt failure and Glasgow Outcome Scale (GOS) at 6 and 12-months. Statistical significance was assessed at p < 0.05. In 402 patients, mean age was 38.0 ± 16.7 years and 315 (78.4%) were male. Forty (10.0%) patients developed PTH, with predominant injuries being subdural hemorrhage (36.4%) and diffuse axonal injury (36.4%). Decompressive hemicraniectomy (DHC) was associated with hydrocephalus (OR 3.62, 95% CI (1.62-8.07), p < 0.01). Eighteen (4.5%) patients had shunt failure and proximal obstruction was most common. Differences in baseline CSF cell count were associated with increased shunt failure. PTH was not associated with worse outcomes at 6 (p = 0.55) or 12 (p = 0.47) months. Hydrocephalus is a frequent sequela in 10.0% of patients, particularly after DHC. Shunt placement and revision procedures are common after severe TBI, within the first 4 months of injury and necessitates early recognition by the clinician.

Applying the Sliding Scale Approach to Quantifying Functional Outcomes Up to Two Years After Severe Traumatic Brain Injury

Outcomes after severe traumatic brain injury (TBI) can be represented by a sliding score that compares actual functional recovery to that predicted by illness severity models. This approach has been applied in clinical trials because of its statistical efficiency and interpretability but has not been used to describe change in functional recovery over time. The objective of this study was to use a sliding scoring system to describe the magnitude of change in Glasgow Outcome Scale Extended (GOSE) score at 6, 12, and 24 months after severe TBI and to compare patients who improved after 6 months to those who did not. This study included consecutive severe TBI patients (Glasgow Coma Scale ≤8; n = 482) from a single center. We grouped patients into four strata based on probability of unfavorable outcome (GOSE = 1-4) using the International Mission on Prognosis and Analysis of Clinical Trials (IMPACT) model, selected a dichotomous GOSE threshold within each stratum, and compared each patient's GOSE to this threshold to calculate a score (GOSE-Sliding Scale [SS]) from -5 to +4 at 6, 12, and 24 months. We compared GOSE-SS at 6 months with GOSE-SS at 12 and 24 months and also compared characteristics of participants who improved after 6 months with characteristics of those who did not using χ2 and t tests. Compared with at 6 months, 40% of patients (n = 74) had improved GOSE-SS at 12 months, and 53% had improved GOSE-SS by 24 months (n = 72). Among those who improved at 12 months, the average magnitude of improvement was 1.7 ± 0.9 and among those who improved at 24 months, the average magnitude of improvement was 1.9 ± 1.0. Those who improved their GOSE-SS score from 6 to 24 months had longer hospital stays (mean-difference = 8.6 days; p = 0.03), longer intensive care unit (ICU) stays (mean-difference = 5.5 days; p = 0.02), and longer ventilator time (mean-difference = 5 days; p = 0.02) than those who worsened. These results support an optimistic long-term outlook for severe TBI patients and emphasize the importance of long-term follow-up in severe TBI survivors.

Cervical intramedullary spinal cavernoma in setting of unresolved myelopathy: A case report

Background:

Spinal cavernous malformations are rare, accounting for approximately 5–12% of all spinal cord vascular lesions. Fortunately, improvements in imaging technologies have made it easier to establish the diagnosis of intramedullary spinal cavernomas (ISCs).

Case Description:

Here, we report the case of a 63-year-old male with an >11-year history of left-sided radiculopathy, ataxia, and quadriparesis. Initially, radiographic findings were interpreted as consistent with spondylotic myelopathy with cord signal changes from the C3-C7 levels. The patient underwent a C3-C7 laminectomy/foraminotomy with instrumentation. It was only after several symptomatic recurrences and repeated magnetic resonance images (MRI) that the diagnosis of a ventrally-located intramedullary lesion, concerning for a cavernoma, at the level C6 was established.

Conclusion:

Early and repeated enhanced MR studies may be required to correctly establish the diagnosis and determine the optimal surgical management of ISCs.

Interaction of admission platelet count with current medications and the risk for chronic subdural recurrence

OBJECTIVE

Chronic subdural hematoma (cSDH) has a reported 10%-24% rate of recurrence after surgery, and prognostic models for recurrence have produced equivocal results. The objective of this study was to leverage a data mining algorithm, chi-square automatic interaction detection (CHAID), which can incorporate continuous, nominal, and binary data into a decision tree, to identify the most robust predictors of repeat surgery for cSDH patients.

METHODS

This was a retrospective cohort study of all patients with SDH from two level 1 trauma centers at a single institution. All patients underwent cSDH evacuation performed by 15 neurosurgeons between 2011 and 2020. The primary outcome was the rate of repeat surgery for recurrent cSDH following the initial evacuation. The authors used CHAID to identify relevant predictors of repeat surgery, including age, sex, comorbidities, postsurgical complications, platelet count prior to the first procedure, midline shift prior to the first procedure, hematoma volume, and preoperative use of anticoagulants, antiplatelets, or statins.

RESULTS

Sixty (13.8%) of 435 study-eligible patients (average age 74.0 years) had a cSDH recurrence. These patients had 2.0 times greater odds of having used anticoagulants. The final CHAID model had an overall accuracy of 87.4% and an area under the curve of 0.76. According to the model, the predictor with the strongest association with cSDH recurrence was admission platelet count. Approximately 26% of patients (n = 23/87) with an admission platelet count < 157 × 109/L had a cSDH recurrence, whereas none of the 44 patients with admission platelets > 313 × 109/L had a recurrence. Approximately 17% of patients in the 157-313 × 109/L platelet group who had used preoperative statins required a second procedure, which was associated with a 2.3 times increased risk for repeat surgery compared to those who had not used statins preoperatively. Among those who had not used preoperative statins, a platelet count ≤ 179 × 109/L on admission for the first procedure was the strongest differentiator for a second surgery (n = 5/22 [23%]), which increased the risk of recurrence by 4.5 times. Among the patients using preoperative statins, the use of anticoagulants was the strongest differentiator for requiring repeat surgery (n = 11/33 [33%]).

CONCLUSIONS

The described model identified platelet count on admission as the most important predictor of repeat cSDH surgery, followed by preoperative statin use and anticoagulant use. Critical cutoffs for platelet count were identified, which future studies should evaluate to determine if they are modifiable or reflective of underlying disease states.

Focal brain oxygen, blood flow, and intracranial pressure measurements in relation to optimal cerebral perfusion pressure

OBJECTIVE

Different paradigms for neurocritical care of traumatic brain injury (TBI) have emerged in conjunction with advanced neuromonitoring technologies and derived metrics. The priority for optimizing these metrics is not currently clear. The goal of this study was to determine whether achieving cerebral perfusion pressure (CPPopt) also improves other metrics like brain oxygenation and brain blood flow.

METHODS

The authors performed a retrospective analysis of high-frequency data from patients with TBI who were treated at a single center and who had partial pressure of brain oxygen (PbtO2) measurements and/or brain blood flow measurements, while also undergoing intracranial pressure (ICP) monitoring. CPPopt was not calculated or targeted during patient care, but was retrospectively computed, as was the difference between the observed CPP and CPPopt.

RESULTS

A total of 22 patients with ICP, PbtO2, and/or brain blood flow monitoring were included in the analysis, and 245.7 days of measurements obtained every second were analyzed including 6,748,866 PbtO2 measurements, 3,296,405 blood flow measurements, and 10,264,770 ICP measurements. The data obtained every second were averaged by minute for analysis. In summative data, PbtO2 measurements peaked near CPPopt and were not improved above CPPopt. Blood flow measurements remained stable near CPPopt, decreased below it, and increased when CPP exceeded CPPopt. ICP decreased linearly with CPP without a specific relationship with CPPopt. In an inverse analysis, the percentage of CPP values at CPPopt, although significantly higher on the favorable side of contemporary treatment thresholds of PbtO2, ICP, and blood flow, was not found to be strongly correlated with the mean values of the physiological measurements obtained every minute (r = 0.27, r = 0.11, and r = 0.47 for ICP, PbtO2, and blood flow, respectively; p < 0.0001).

CONCLUSIONS

Although CPPopt was not targeted in the patients in this study, CPPopt was a physiologically significant value based on concurrent measurements of PbtO2 and blood flow. In summative data, achievement of CPPopt was associated with optimized PbtO2 and blood flow. Conversely, the correlation between achievement of CPPopt and the mean measurement value was not strong, strengthening the significance of CPPopt. In individual patients, achieving CPPopt is not always associated with optimal PbtO2 or blood flow. Further research should explore these relationships in treatment paradigms that specifically target CPPopt. These data do not support the premise that targeting and achieving CPPopt obviates the need for concurrent PbtO2 and blood flow monitoring. Although these data suggest that targeting CPPopt may be an appropriate initial treatment strategy, they do not provide evidence that CPPopt should be targeted with highest priority.

Automated Preoperative and Postoperative Volume Estimates Risk of Retreatment in Chronic Subdural Hematoma: A Retrospective, Multicenter Study

BACKGROUND AND OBJECTIVES

Several neurosurgical pathologies, ranging from glioblastoma to hemorrhagic stroke, use volume thresholds to guide treatment decisions. For chronic subdural hematoma (cSDH), with a risk of retreatment of 10%-30%, the relationship between preoperative and postoperative cSDH volume and retreatment is not well understood. We investigated the potential link between preoperative and postoperative cSDH volumes and retreatment.

METHODS

We performed a retrospective chart review of patients operated for unilateral cSDH from 4 level 1 trauma centers, February 2009-August 2021. We used a 3-dimensional deep learning, automated segmentation pipeline to calculate preoperative and postoperative cSDH volumes. To identify volume thresholds, we constructed a receiver operating curve with preoperative and postoperative volumes to predict cSDH retreatment rates and selected the threshold with the highest Youden index. Then, we developed a light gradient boosting machine to predict the risk of cSDH recurrence.

RESULTS

We identified 538 patients with unilateral cSDH, of whom 62 (12%) underwent surgical retreatment within 6 months of the index surgery. cSDH retreatment was associated with higher preoperative (122 vs 103 mL; P < .001) and postoperative (62 vs 35 mL; P < .001) volumes. Patients with >140 mL preoperative volume had nearly triple the risk of cSDH recurrence compared with those below 140 mL, while a postoperative volume >46 mL led to an increased risk for retreatment (22% vs 6%; P < .001). On multivariate modeling, our model had an area under the receiver operating curve of 0.76 (95% CI: 0.60-0.93) for predicting retreatment. The most important features were preoperative and postoperative volume, platelet count, and age.

CONCLUSION

Larger preoperative and postoperative cSDH volumes increase the risk of retreatment. Volume thresholds may allow identification of patients at high risk of cSDH retreatment who would benefit from adjunct treatments. Machine learning algorithm can quickly provide accurate estimates of preoperative and postoperative volumes.

Isolated traumatic occipital condyle fractures: Is external cervical orthosis even necessary?

Background:

Occipital condyle fractures (OCFs) have been reported in up to 4–16% of individuals suffering cervical spine trauma. The current management of OCF fractures relies on a rigid cervical collar for 6 weeks or longer. Here, we calculated the rate of acute and delayed surgical intervention (occipitocervical fusion) for patients with isolated OCF who were managed with a cervical collar over a 10-year period at a single institution.

Methods:

This was a retrospective analysis performed on all patients admitted to a Level 1 Trauma Center between 2008 and 2018 who suffered traumatic isolated OCF managed with an external rigid cervical orthosis. Radiographic imaging was reviewed by several board-certified neuroradiologists. Demographic and clinical data were collected including need for occipitocervical fusion within 12 months after trauma.

Results:

The incidence of isolated OCF was 4% (60/1536) for those patients admitted with cervical spine fractures. They averaged 49 years of age, and 58% were male falls accounted for the mechanism of injury in 47% of patients. Classification of OCF was most commonly classified in 47% as type I Anderson and Montesano fractures. Of the 60 patients who suffered isolated OCF that was managed with external cervical orthosis, 0% required occipitocervical fusion within 12 months posttrauma. About 90% were discharged, while the remaining 10% sustained traumatic brain/orthopedic injury that limited an accurate neurological assessment.

Conclusion:

Here, we documented a 4% incidence of isolated OCF in our cervical trauma population, a rate which is comparable to that found in the literature year. Most notably, we documented a 0% incidence for requiring delayed occipital-cervical fusions.

Successful management of an intraluminal superior sagittal sinus meningioma causing elevated intracranial pressure using gamma knife radiosurgery in subacute setting: A case report

Background:

Gamma Knife stereotactic radiosurgery (GKRS) facilitates precisely focused radiation to an intracranial target while minimizing substantial off-target radiation in the surrounding normal tissue. Meningiomas attached to or invading the superior sagittal sinus may result in sinus occlusion and are often impossible to completely resect safely. The authors describe successful management of a patient with a meningioma located completely inside the posterior aspect of the superior sagittal sinus.

Case Description:

A 46-year-old woman presented to the emergency department with progressive generalized headaches accompanied by worsening vision. The patient underwent a diagnostic brain magnetic resonance imaging which showed a solitary a 7 × 6 × 10 mm homogeneously contrast-enhancing lesion within the lumen of the posterior aspect of superior sagittal sinus without ventricular enlargement or peritumoral edema. The lesion was thought to be a meningioma radiographically. To evaluate the suspected increased intracranial pressure, a lumbar puncture was subsequently performed and demonstrated an opening pressure of 30 cm H2O. After drainage of 40 cc of CSF, the spinal closing pressure was 9 cm H2O. After failure of conservative management with acetazolamide, and determination of surgical inoperability due to the critical intraluminal location of the mass lesion, the patient underwent Gamma Knife radiosurgery. The 0.36 cc tumor was treated as an outpatient in the Perfexion® model Gamma Knife with a highly conformal and selective plan that enclosed the 3D geometry of the tumor with a minimal margin tumor dose of 14 gy at the 50% isodose. Three months after GKRS, the patient reported continued reduction in the frequency and severity of both her headaches and her visual disturbance. Ophthalmological consultation noted progressive resolution of her optic disc edema confirmed by formal optical coherence tomography. The patient is now 3 years out from GKRS with complete resolution of headache symptoms along with persistent reduction in tumor size (3 × 1 × 4 mm) on serial period imaging and resolution of papilledema.

Conclusion:

Tumors located in such critical anatomic regions, as in our patient, should be considered for primary GKRS when the risks of biopsy or removal are too high. GKRS was able to provide great radiographic and clinical result in an intricately located meningioma.