Logo: to the web site of Uppsala University

Background: A major challenge in management of traumatic brain injury (TBI) is to assess the heterogeneity of TBI pathology and outcome prediction. A reliable outcome prediction would have both great value for the healthcare provider, but also for the patients and their relatives. A well-known prediction model is the International Mission for Prognosis and Analysis of Clinical Trials (IMPACT) prognostic calculator. The aim of this study was to externally validate all three modules of the IMPACT calculator on TBI patients admitted to Uppsala University hospital (UUH).

Method: TBI patients admitted to UUH are continuously enrolled into the Uppsala neurointensive care unit (NICU) TBI Uppsala Clinical Research (UCR) quality register. The register contains both clinical and demographic data, radiological evaluations, and outcome assessments based on the extended Glasgow outcome scale extended (GOSE) performed at 6 months to 1 year. In this study, we included 635 patients with severe TBI admitted during 2008–2020. We used IMPACT core parameters: age, motor score, and pupillary reaction.

Results: The patients had a median age of 56 (range 18–93), 142 female and 478 male. Using the IMPACT Core model to predict outcome resulted in an AUC of 0.85 for mortality and 0.79 for unfavorable outcome. The CT module did not increase AUC for mortality and slightly decreased AUC for unfavorable outcome to 0.78. However, the lab module increased AUC for mortality to 0.89 but slightly decreased for unfavorable outcome to 0.76. Comparing the predicted risk to actual outcomes, we found that all three models correctly predicted low risk of mortality in the surviving group of GOSE 2–8. However, it produced a greater variance of predicted risk in the GOSE 1 group, denoting general underprediction of risk. Regarding unfavorable outcome, all models once again underestimated the risk in the GOSE 3–4 groups, but correctly predicts low risk in GOSE 5–8.

Conclusions: The results of our study are in line with previous findings from centers with modern TBI care using the IMPACT model, in that the model provides adequate prediction for mortality and unfavorable outcome. However, it should be noted that the prediction is limited to 6 months outcome and not longer time interval.

Outcome after traumatic brain injury (TBI) is typically assessed using the Glasgow outcome scale extended (GOSE) with levels from 1 (death) to 8 (upper good recovery). Outcome prediction has classically been dichotomized into either dead/alive or favorable/unfavorable outcome. Binary outcome prediction models limit the possibility of detecting subtle yet significant improvements. We set out to explore different machine learning methods with the purpose of mapping their predictions to the full 8 grade scale GOSE following TBI. The models were set up using the variables: age, GCS-motor score, pupillary reaction, and Marshall CT score. For model setup and internal validation, a total of 866 patients could be included. For external validation, a cohort of 369 patients were included from Leuven, Belgium, and a cohort of 573 patients from the US multi-center ProTECT III study. Our findings indicate that proportional odds logistic regression (POLR), random forest regression, and a neural network model achieved accuracy values of 0.3–0.35 when applied to internal data, compared to the random baseline which is 0.125 for eight categories. The models demonstrated satisfactory performance during external validation in the data from Leuven, however, their performance were not satisfactory when applied to the ProTECT III dataset.

This study aimed to evaluate the predictive value and clinical impact of a clinically implemented artificial neural network software model. The software detects intracranial hemorrhage (ICH) from head computed tomography (CT) scans and artificial intelligence (AI)-identified positive cases are then annotated in the work list for early radiologist evaluation. The index test was AI detection by the program Zebra Medical Vision-HealthICH+. Radiologist-confirmed ICH was the reference standard. The study compared whether time benefits from using the AI model led to faster escalation of patient care or surgery within the first 24 h. A total of 2,306 patients were evaluated by the software, and 288 AI-positive cases were included. The AI tool had a positive predictive value of 0.823. There was, however, no significant time reduction when comparing the patients who required escalation of care and those who did not. There was also no significant time reduction in those who required acute surgery compared with those who did not. Among the individual patients with reduced time delay, no cases with evident clinical benefit were identified. Although the clinically implemented AI-based decision support system showed adequate predictive value in identifying ICH, there was no significant clinical benefit for the patients in our setting. While AI-assisted detection of ICH shows great promise from a technical perspective, there remains a need to evaluate the clinical impact and perform external validation across different settings.