Decision tree analysis in subarachnoid hemorrhage: prediction of outcome parameters during the course of aneurysmal subarachnoid hemorrhage using decision tree analysis

Comparison of an Ensemble of Machine Learning Models and the BERT Language Model for Analysis of Text Descriptions of Brain CT Reports to Determine the Presence of Intracranial Hemorrhage

The aim of this study is to train and test an ensemble of machine learning models, as well as to compare its performance with the BERT language model pre-trained on medical data to perform simple binary classification, i.e., determine the presence/absence of the signs of intracranial hemorrhage (ICH) in brain CT reports.

Materials and Methods

Seven machine learning algorithms and three text vectorization techniques were selected as models to solve the binary classification problem. These models were trained on textual data represented by 3980 brain CT reports from 56 inpatient medical facilities in Moscow. The study utilized three text vectorization techniques: bag of words, TF-IDF, and word2vec. The resulting data were then processed by the following machine learning algorithms: decision tree, random forest, logistic regression, nearest neighbors, support vector machines, Catboost, and XGboost. Data analysis and pre-processing were performed using NLTK (Natural Language Toolkit, version 3.6.5), libraries for character-based and statistical processing of natural language, and Scikit-learn (version 0.24.2), a library for machine learning containing tools to tackle classification challenges. MedRuBertTiny2 was taken as a BERT transformer model pre-trained on medical data.

Results

Based on the training and testing outcomes from seven machine learning algorithms, the authors selected three algorithms that yielded the highest metrics (i.e. sensitivity and specificity): CatBoost, logistic regression, and nearest neighbors. The highest metrics were achieved by the bag of words technique. These algorithms were assembled into an ensemble using the stacking technique. The sensitivity and specificity for the validation dataset separated from the original sample were 0.93 and 0.90, respectively. Next, the ensemble and the BERT model were trained on an independent dataset containing 9393 textual radiology reports also divided into training and test sets. Once the ensemble was tested on this dataset, the resulting sensitivity and specificity were 0.92 and 0.90, respectively. The BERT model tested on these data demonstrated a sensitivity of 0.97 and a specificity of 0.90.

Conclusion

When analyzing textual reports of brain CT scans with signs of intracranial hemorrhage, the trained ensemble demonstrated high accuracy metrics. Still, manual quality control of the results is required during its application. The pre-trained BERT transformer model, additionally trained on diagnostic textual reports, demonstrated higher accuracy metrics (p<0.05). The results show promise in terms of finding specific values for both binary classification task and in-depth analysis of unstructured medical information.

Classification and Regression Tree Predictive Model for Acute Kidney Injury in Traumatic Brain Injury Patients

Background

Acute kidney injury (AKI) is prevalent in hospitalized patients with traumatic brain injury (TBI), and increases the risk of poor outcomes. We designed this study to develop a visual and convenient decision-tree-based model for predicting AKI in TBI patients.

Methods

A total of 376 patients admitted to the emergency department of the West China Hospital for TBI between January 2015 and June 2019 were included. Demographic information, vital signs on admission, laboratory test results, radiological signs, surgical options, and medications were recorded as variables. AKI was confirmed since the second day after admission, based on the Kidney Disease Improving Global Outcomes criteria. We constructed two predictive models for AKI using least absolute shrinkage and selection operator (LASSO) regression and classification and regression tree (CART), respectively. Receiver operating characteristic (ROC) curves of these two predictive models were drawn, and the area under the ROC curve (AUC) was calculated to compare their predictive accuracy.

Results

The incidence of AKI on the second day after admission was 10.4% among patients with TBI. Lasso regression identified five potent predictive factors for AKI: glucose, serum creatinine, cystatin C, serum uric acid, and fresh frozen plasma transfusions. The CART analysis showed that glucose, serum uric acid, and cystatin C ranked among the top three in terms of the feature importance of the decision tree model. The AUC value of the decision-tree predictive model was 0.892, which was higher than the 0.854 of the LASSO regression model, although the difference was not statistically significant.

Conclusion

The decision tree model is valuable for predicting AKI among patients with TBI. This tree-based flowchart is convenient for physicians to identify patients with TBI who are at high risk of AKI and prompts them to develop suitable therapeutic strategies.

Clinical prediction models for aneurysmal subarachnoid hemorrhage: a systematic review update

BACKGROUND

A systematic review of clinical prediction models for aneurysmal subarachnoid hemorrhage (aSAH) reported in 2011 noted that clinical prediction models for aSAH were developed using poor methods and were not externally validated. This study aimed to update the above review to guide the future development of predictive models in aSAH.

METHODS

We systematically searched Embase and MEDLINE databases (January 2010 to February 2022) for articles that reported the development of a clinical prediction model to predict functional outcomes in aSAH. Our reviews are based on the items included in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses statement (PRISMA) checklist, and on data abstracted from each study in accord with the Checklist for critical Appraisal and data extraction for systematic Reviews of prediction Modelling Studies (CHARMS) 2014 checklist. Bias and applicability were assessed using the Prediction model Risk Of Bias Assessment Tool (PROBAST).

RESULTS

We reviewed data on 30 466 patients contributing to 29 prediction models abstracted from 22 studies identified from an initial search of 7858 studies. Most models were developed using logistic regression (n=20) or machine learning (n=9) with prognostic variables selected through a range of methods. Age (n=13), World Federation of Neurological Surgeons (WFNS) grade (n=11), hypertension (n=6), aneurysm size (n=5), Fisher grade (n=12), Hunt and Hess score (n=5), and Glasgow Coma Scale (n=8) were the variables most frequently included in the reported models. External validation was performed in only four studies. All but one model had a high or unclear risk of bias due to poor performance or lack of validation.

CONCLUSION

Externally validated models for the prediction of functional outcome in aSAH patients have now become available. However, most of them still have a high risk of bias.

Predicting the formation of mixed pattern hemorrhages in ruptured middle cerebral artery aneurysms based on a decision tree model: A multicenter study

OBJECTIVE

Mixed-pattern hemorrhages (MPH) commonly occur in ruptured middle cerebral artery (MCA) aneurysms and are associated with poor clinical outcomes. This study aimed to predict the formation of MPH in a multicenter database of MCA aneurysms using a decision tree model.

METHODS

We retrospectively reviewed patients with ruptured MCA aneurysms between January 2009 and June 2020. The MPH was defined as subarachnoid hemorrhages with intracranial hematomas and/or intraventricular hemorrhages and/or subdural hematomas. Univariate and multivariate logistic regression analyses were used to explore the prediction factors of the formation of MPH. Based on these prediction factors, a decision tree model was developed to predict the formation of MPH. Additional independent datasets were used for external validation.

RESULTS

We enrolled 436 patients with ruptured MCA aneurysms detected by computed tomography angiography; 285 patients had MPH (65.4%). A multivariate logistic regression analysis showed that age, aneurysm size, multiple aneurysms, and the presence of a daughter dome were the independent prediction factors of the formation of MPH. The areas under receiver operating characteristic curves of the decision tree model in the training, internal, and external validation cohorts were 0.951, 0.927, and 0.901, respectively.

CONCLUSION

Age, aneurysm size, the presence of a daughter dome, and multiple aneurysms were the independent prediction factors of the formation of MPH. The decision tree model is a useful visual triage tool to predict the formation of MPH that could facilitate the management of unruptured aneurysms in routine clinical work.

Cortical Spreading Depolarization and Delayed Cerebral Ischemia; Rethinking Secondary Neurological Injury in Subarachnoid Hemorrhage

Poor outcomes in Subarachnoid Hemorrhage (SAH) are in part due to a unique form of secondary neurological injury known as Delayed Cerebral Ischemia (DCI). DCI is characterized by new neurological insults that continue to occur beyond 72 h after the onset of the hemorrhage. Historically, it was thought to be a consequence of hypoperfusion in the setting of vasospasm. However, DCI was found to occur even in the absence of radiographic evidence of vasospasm. More recent evidence indicates that catastrophic ionic disruptions known as Cortical Spreading Depolarizations (CSD) may be the culprits of DCI. CSDs occur in otherwise healthy brain tissue even without demonstrable vasospasm. Furthermore, CSDs often trigger a complex interplay of neuroinflammation, microthrombi formation, and vasoconstriction. CSDs may therefore represent measurable and modifiable prognostic factors in the prevention and treatment of DCI. Although Ketamine and Nimodipine have shown promise in the treatment and prevention of CSDs in SAH, further research is needed to determine the therapeutic potential of these as well as other agents.

A predictive model based on random forest for shoulder-hand syndrome

Objectives

The shoulder-hand syndrome (SHS) severely impedes the function recovery process of patients after stroke. It is incapable to identify the factors at high risk for its occurrence, and there is no effective treatment. This study intends to apply the random forest (RF) algorithm in ensemble learning to establish a predictive model for the occurrence of SHS after stroke, aiming to identify high-risk SHS in the first-stroke onset population and discuss possible therapeutic methods.

Methods

We retrospectively studied all the first-onset stroke patients with one-side hemiplegia, then 36 patients that met the criteria were included. The patients’ data concerning a wide spectrum of demographic, clinical, and laboratory data were analyzed. RF algorithms were built to predict the SHS occurrence, and the model’s reliability was measured with a confusion matrix and the area under the receiver operating curves (ROC).

Results

A binary classification model was trained based on 25 handpicked features. The area under the ROC curve of the prediction model was 0.8 and the out-of-bag accuracy rate was 72.73%. The confusion matrix indicated a sensitivity of 0.8 and a specificity of 0.5, respectively. And the feature importance scored the weights (top 3 from large to small) in the classification were D-dimer, C-reactive protein, and hemoglobin.

Conclusion

A reliable predictive model can be established based on post-stroke patients’ demographic, clinical, and laboratory data. Combining the results of RF and traditional statistical methods, our model found that D-dimer, CRP, and hemoglobin affected the occurrence of the SHS after stroke in a relatively small sample of data with tightly controlled inclusion criteria.

VPS dependency after aneurysmal subarachnoid haemorrhage and influence of admission hyperglycaemia

Introduction

Hydrocephalus after aneurysmal subarachnoid haemorrhage (aSAH) is a common complication which may lead to insertion of a ventriculoperitoneal shunt (VPS). Our aim is to evaluate a possible influence of specific clinical and biochemical factors on VPS dependency with special emphasis on hyperglycaemia on admission.

Patients and methods

Retrospective analysis of a monocentric database of aSAH patients. Using univariable and multivariable logistic regression analysis we evaluated factors influencing VPS dependency, with a special focus on hyperglycaemia on blood sample within 24 h of admission, dichotomised at 126 mg/dl. Factors evaluated in the univariable analysis were age, sex, known diabetes, Hunt and Hess grade, Barrow Neurological Institute scale, treatment modality, extra-ventricular drain (EVD) insertion, complications (rebleeding, vasospasm, infarction, decompressive craniectomy, ventriculitis), outcome variables and laboratory parameters (glucose, C-reactive protein, procalcitonin).

Results

We included 510 consecutive patients treated with acute aSAH requiring a VPS (mean age 58.2 years, 66% were female). An EVD was inserted in 387 (75.9%) patients. In the univariable analysis, VPS dependency was associated with hyperglycaemia on admission (OR 2.56, 95%CI 1.58-4.14, p < 0.001). In the multivariable regression analysis after stepwise backward regression, factors associated with VPS dependency were hyperglycaemia >126 mg/dl on admission (OR 1.93, 95%CI 1.13-3.30, p = 0.02), ventriculitis (OR 2.33, 95%CI 1.33-4.04, p = 0.003), Hunt and Hess grade (overall p-value 0.02) and decompressive craniectomy (OR 2.68, 95%CI 1.55-4.64, p < 0.001).

Conclusion

Hyperglycaemia on admission was associated with an increased probability of VPS placement. If confirmed, this finding might facilitate treatment of these patients by accelerating insertion of a permanent draining system.

Intracranial Hemorrhage Detection Using Parallel Deep Convolutional Models and Boosting Mechanism

Intracranial hemorrhage (ICH) can lead to death or disability, which requires immediate action from radiologists. Due to the heavy workload, less experienced staff, and the complexity of subtle hemorrhages, a more intelligent and automated system is necessary to detect ICH. In literature, many artificial-intelligence-based methods are proposed. However, they are less accurate for ICH detection and subtype classification. Therefore, in this paper, we present a new methodology to improve the detection and subtype classification of ICH based on two parallel paths and a boosting technique. The first path employs the architecture of ResNet101-V2 to extract potential features from windowed slices, whereas Inception-V4 captures significant spatial information in the second path. Afterwards, the detection and subtype classification of ICH is performed by the light gradient boosting machine (LGBM) using the outputs of ResNet101-V2 and Inception-V4. Thus, the combined solution, known as ResNet101-V2, Inception-V4, and LGBM (Res-Inc-LGBM), is trained and tested over the brain computed tomography (CT) scans of CQ500 and Radiological Society of North America (RSNA) datasets. The experimental results state that the proposed solution efficiently obtains 97.7% accuracy, 96.5% sensitivity, and 97.4% F1 score using the RSNA dataset. Moreover, the proposed Res-Inc-LGBM outperforms the standard benchmarks for the detection and subtype classification of ICH regarding the accuracy, sensitivity, and F1 score. The results prove the significance of the proposed solution for its real-time application.

External Validation of the HATCH (Hemorrhage, Age, Treatment, Clinical State, Hydrocephalus) Score for Prediction of Functional Outcome After Subarachnoid Hemorrhage

BACKGROUND

The Hemorrhage, Age, Treatment, Clinical State, Hydrocephalus (HATCH) Score has previously shown to predict functional outcome in aneurysmal subarachnoid hemorrhage (aSAH).

OBJECTIVE

To validate the HATCH score.

METHODS

This is a pooled cohort study including prospective collected data on 761 patients with aSAH from 4 different hospitals. The HATCH score for prediction of functional outcome was validated using calibration and discrimination analysis (area under the curve). HATCH score model performance was compared with the World Federation of Neurosurgical Societies and Barrow Neurological Institute score.

RESULTS

At the follow-up of at least 6 months, favorable (Glasgow Outcome Score 4-5) and unfavorable functional outcomes (Glasgow Outcome Score 1-3) were observed in 512 (73%) and 189 (27%) patients, respectively. A higher HATCH score was associated with an increased risk of unfavorable outcome with a score of 1 showing a risk of 1.3% and a score of 12 yielding a risk of 67%. External validation showed a calibration intercept of -0.07 and slope of 0.60 with a Brier score of 0.157 indicating good model calibration and accuracy. With an area under the curve of 0.81 (95% CI 0.77-0.84), the HATCH score demonstrated superior discriminative ability to detect favorable outcome at follow-up compared with the World Federation of Neurosurgical Societies and Barrow Neurological Institute score with 0.72 (95% CI 0.67-0.75) and 0.63 (95% CI 0.59-0.68), respectively.

CONCLUSION

This multicenter external validation analysis confirms the HATCH score to be a strong independent predictor for functional outcome. Its incorporation into daily practice may be of benefit for goal-directed patient care in aSAH.

Brain Oxygen-Directed Management of Aneurysmal Subarachnoid Hemorrhage. Temporal Patterns of Cerebral Ischemia During Acute Brain Attack, Early Brain Injury, and Territorial Sonographic Vasospasm

BACKGROUND

Neurocritical management of aneurysmal subarachnoid hemorrhage focuses on delayed cerebral ischemia (DCI) after aneurysm repair.

METHODS

This study conceptualizes the pathophysiology of cerebral ischemia and its management using a brain oxygen-directed protocol (intracranial pressure [ICP] control, eubaric hyperoxia, hemodynamic therapy, arterial vasodilation, and neuroprotection) in patients with subarachnoid hemorrhage, undergoing aneurysm clipping (n = 40).

RESULTS

The brain oxygen-directed protocol reduced Lbo₂ (Pbto₂ [partial pressure of brain tissue oxygen] <20 mm Hg) from 67% to 15% during acute brain attack (<24 hours of ictus), by increasing Pbto₂ from 11.31 ± 9.34 to 27.85 ± 6.76 (P < 0.0001) and then to 29.09 ± 17.88 within 72 hours. Day-after-bleed, Fio₂ change, ICP, hemoglobin, and oxygen saturation were predictors for Pbto₂ during early brain injury. Transcranial Doppler ultrasonography velocities (>20 cm/second) increased at day 2. During DCI caused by territorial sonographic vasospasm (TSV), middle cerebral artery mean velocity (V_m) increased from 45.00 ± 15.12 to 80.37 ± 38.33/second by day 4 with concomitant Pbto₂ reduction from 29.09 ± 17.88 to 22.66 ± 8.19. Peak TSV (days 7-12) coincided with decline in Pbto₂. Nicardipine mitigated Lbo₂ during peak TSV, in contrast to nimodipine, with survival benefit (P < 0.01). Intravenous and cisternal nicardipine combination had survival benefit (Cramer Φ = 0.43 and 0.327; G² = 28.32; P < 0.001). This study identifies 4 zones of Lbo₂ during survival benefit (Cramer Φ = 0.43 and 0.3) TSV, uncompensated; global cerebral ischemia, compensated, and normal Pbto₂. Admission Glasgow Coma Scale score (not increased ICP) was predictive of low Pbto₂ (β = 0.812, R² = 0.661, F_1,30 = 58.41; P < 0.0001) during early brain injury. Coma was the only credible predictor for mortality (odds ratio, 7.33/>4.8∗; χ² = 7.556; confidence interval, 1.70-31.54; P < 0.01) followed by basilar aneurysm, poor grade, high ICP and Lbo₂ during TSV. Global cerebral ischemia occurs immediately after the ictus, persisting in 30% of patients despite the high therapeutic intensity level, superimposed by DCI during TSV.

CONCLUSIONS

We propose implications for clinical practice and patient management to minimize cerebral ischemia.

Task-Based Visual Interactive Modeling: Decision Trees and Rule-Based Classifiers

Visual analytics enables the coupling of machine learning models and humans in a tightly integrated workflow, addressing various analysis tasks. Each task poses distinct demands to analysts and decision-makers. In this survey, we focus on one canonical technique for rule-based classification, namely decision tree classifiers. We provide an overview of available visualizations for decision trees with a focus on how visualizations differ with respect to 16 tasks. Further, we investigate the types of visual designs employed, and the quality measures presented. We find that (i) interactive visual analytics systems for classifier development offer a variety of visual designs, (ii) utilization tasks are sparsely covered, (iii) beyond classifier development, node-link diagrams are omnipresent, (iv) even systems designed for machine learning experts rarely feature visual representations of quality measures other than accuracy. In conclusion, we see a potential for integrating algorithmic techniques, mathematical quality measures, and tailored interactive visualizations to enable human experts to utilize their knowledge more effectively.

Association between acute kidney injury and long-term mortality in patients with aneurysmal subarachnoid hemorrhage: A retrospective study

Background

Though acute kidney injury (AKI) in the context of aneurysmal subarachnoid hemorrhage (aSAH) worsens short-term outcomes, its impact on long-term survival is unknown.

Aim

We aimed to evaluate the association between long-term mortality and AKI during hospitalization for aSAH.

Methods

This was a retrospective study of patients who survived &gt;12 months after aSAH. All patients were evaluated at West China Hospital, Sichuan University, between December 2013 and June 2019. The minimum follow-up time was over 1 year. the maximum follow-up time was about 7.3 years. AKI was defined by the KDIGO (The Kidney Disease Improving Global Outcomes) guidelines, which stratifies patients into three stages of severity. The primary outcome was long-term mortality, which was analyzed with Kaplan-Meier curves and Cox proportional hazards models.

Results

During this study period, 238 (9.2%) patients had AKI among 2,592 patients with aSAH. We confirmed that AKI during care for aSAH significantly increased long-term mortality (median 4.3 years of follow-up) and that risk increased with the severity of the kidney failure, with an adjusted hazard ratio (HR) of 2.08 (95% CI 1.49–2.89) for stage 1 AKI, 2.15 (95% CI 1.05–4.43) for stage 2 AKI, and 2.66 (95% CI 1.08–6.53) for stage 3 AKI compared with patients without AKI. Among patients with an AKI episode, those with renal recovery still had increased long-term mortality (HR 1.96; 95% CI 1.40–2.74) compared with patients without AKI but had better long-term outcomes than those without renal recovery (HR 0.51, 95% CI 0.27–0.97).

Conclusions

Among 12-month survivors of aSAH, AKI during their initial hospitalization for aSAH was associated with increased long-term mortality, even for patients who had normal renal function at the time of hospital discharge. Longer, multidisciplinary post-discharge follow-up may be warranted for these patients.

An accurate prognostic prediction for aneurysmal subarachnoid hemorrhage dedicated to patients after endovascular treatment

Background

Endovascular treatment for aneurysmal subarachnoid hemorrhage (aSAH) has high fatality and permanent disability rates. It remains unclear how the prognosis is determined by the complex interaction between clinical severity and aneurysm characteristics.

Objective

This study aimed to design an accurate prognostic prediction model for aSAH patients after endovascular treatment and elucidate the interaction between clinical severity and aneurysm characteristics.

Methods

We used a clinically homogeneous data set with 1029 aSAH patients who received endovascular treatment to develop prognostic models. Aneurysm characteristics were measured by variables, such as aneurysm size, neck size, and dome-to-neck ratio, while clinical severity on admission was measured by both comorbidities and neurological condition. In total, 18 clinical variables were used for prognostic prediction. Considering the imbalance between the favorable and the poor outcomes in this clinical population, both ensemble learning and deep reinforcement learning approaches were used for prediction.

Results

The random forest (RF) model was selected as the best approach for the prognostic prediction for all patients and also for patients with good-grade aSAH. Using an independent test data set, the model made accurate predictions (AUC = 0.869 ± 0.036, sensitivity = 0.709 ± 0.087, specificity = 0.805 ± 0.034) with the clinical severity on admission as a leading contributor to the prediction. For patients with good-grade aSAH, the RF model performed the best (AUC = 0.805 ± 0.034, sensitivity = 0.620 ± 0.172, specificity = 0.696 ± 0.043) with aneurysm characteristics as leading contributors. The classic scoring systems failed in this patient group (AUC < 0.600; sensitivity = 0.000, specificity = 1.000).

Conclusion

The proposed prognostic prediction model outperformed the classic scoring systems for patients with aSAH after endovascular treatment, especially when the classic scoring systems failed to make any informative prediction for patients with good-grade aSAH, who constitute the majority group (79%) of this clinical population.

Could outcomes of intracranial aneurysms be better predict using serum creatinine and glomerular filtration rate?

Purpose:

To analyze the role of serum creatinine levels as a biomarker of intracranial aneurysm outcomes.

Methods:

This is a prospective analysis of outcomes of patients with intracranial aneurysm. One hundred forty-seven patients with serum creatinine at admission and 6 months follow up were included. Linear and logistic regressions were used to analyze the data. Modified Rankin scale (mRS) was used to assess outcome.

Results:

Creatinine level was not directly related to aneurysm outcome nor aneurysm rupture (p > 0.05). However, patients with a glomerular filtration rate (GFR) lower than 72.50 mL·min^–1 had an odds ratio (OR) of 3.049 (p = 0.006) for worse outcome. Similarly, aneurysm rupture had an OR of 2.957 (p = 0.014) for worse outcomes. Stepwise selection model selected 4 variables for outcomes prediction: serum creatinine, sex, hypertension and treatment. Hypertensive patients had, on average, an increase in 0.588 in mRS (p = 0.022), while treatment with microsurgery had a decrease in 0.555 (p = 0.038).

Conclusions:

Patients with higher GFR had better outcomes after 6 months. Patients with higher GFR had better outcomes after 6 months. Creatinine presented an indirect role in GFR values and should be included in models for outcome prediction.

A Comparison of LASSO Regression and Tree-Based Models for Delayed Cerebral Ischemia in Elderly Patients With Subarachnoid Hemorrhage

Backgrounds

As a most widely used machine learning method, tree-based algorithms have not been applied to predict delayed cerebral ischemia (DCI) in elderly patients with aneurysmal subarachnoid hemorrhage (aSAH). Hence, this study aims to develop the conventional regression and tree-based models and determine which model has better prediction performance for DCI development in hospitalized elderly patients after aSAH.

Methods

This was a multicenter, retrospective, observational cohort study analyzing elderly patients with aSAH aged 60 years and older. We randomly divided the multicentral data into model training and validation cohort in a ratio of 70–30%. One conventional regression and tree-based model, such as least absolute shrinkage and selection operator (LASSO), decision tree (DT), random forest (RF), and eXtreme Gradient Boosting (XGBoost), was developed. Accuracy, sensitivity, specificity, area under the precision-recall curve (AUC-PR), and area under the receiver operating characteristic curve (AUC-ROC) with 95% CI were employed to evaluate the model prediction performance. A DeLong test was conducted to calculate the statistical differences among models. Finally, we figured the importance weight of each feature to visualize the contribution on DCI.

Results

There were 111 and 42 patients in the model training and validation cohorts, and 53 cases developed DCI. According to AUC-ROC value in the model internal validation, DT of 0.836 (95% CI: 0.747–0.926, p = 0.15), RF of 1 (95% CI: 1–1, p < 0.05), and XGBoost of 0.931 (95% CI: 0.885–0.978, p = 0.01) outperformed LASSO of 0.793 (95% CI: 0.692–0.893). However, the LASSO scored a highest AUC-ROC value of 0.894 (95% CI: 0.8–0.989) than DT of 0.764 (95% CI: 0.6–0.928, p = 0.05), RF of 0.821 (95% CI: 0.683–0.959, p = 0.27), and XGBoost of 0.865 (95% CI: 0.751–0.979, p = 0.69) in independent external validation. Moreover, the LASSO had a highest AUC-PR value of 0.681 than DT of 0.615, RF of 0.667, and XGBoost of 0.622 in external validation. In addition, we found that CT values of subarachnoid clots, aneurysm therapy, and white blood cell counts were the most important features for DCI in elderly patients with aSAH.

Conclusions

The LASSO had a superior prediction power than tree-based models in external validation. As a result, we recommend the conventional LASSO regression model to predict DCI in elderly patients with aSAH.

External Validation of a Neural Network Model in Aneurysmal Subarachnoid Hemorrhage: A Comparison With Conventional Logistic Regression Models

BACKGROUND

Interest in machine learning (ML)-based predictive modeling has led to the development of models predicting outcomes after aneurysmal subarachnoid hemorrhage (aSAH), including the Nijmegen acute subarachnoid hemorrhage calculator (Nutshell). Generalizability of such models to external data remains unclear.

OBJECTIVE

To externally validate the performance of the Nutshell tool while comparing it with the conventional Subarachnoid Hemorrhage International Trialists (SAHIT) models and to review the ML literature on outcome prediction after aSAH and aneurysm treatment.

METHODS

A prospectively maintained database of patients with aSAH presenting consecutively to our institution in the 2013 to 2018 period was used. The web-based Nutshell and SAHIT calculators were used to derive the risks of poor long-term (12-18 months) outcomes and 30-day mortality. Discrimination was evaluated using the area under the curve (AUC), and calibration was investigated using calibration plots. The literature on relevant ML models was surveyed for a synopsis.

RESULTS

In 269 patients with aSAH, the SAHIT models outperformed the Nutshell tool (AUC: 0.786 vs 0.689, P = .025) in predicting long-term functional outcomes. A logistic regression model of the Nutshell variables derived from our data achieved adequate discrimination (AUC = 0.759) of poor outcomes. The SAHIT models outperformed the Nutshell tool in predicting 30-day mortality (AUC: 0.810 vs 0.636, P < .001). Calibration properties were more favorable for the SAHIT models. Most published aneurysm-related ML-based outcome models lack external validation and usable testing platforms.

CONCLUSION

The Nutshell tool demonstrated limited performance on external validation in comparison with the SAHIT models. External validation and the dissemination of testing platforms for ML models must be emphasized.

OUP accepted manuscript

ICU Cockpit: a secure, fast, and scalable platform for collecting multimodal waveform data, online and historical data visualization, and online validation of algorithms in the intensive care unit. We present a network of software services that continuously stream waveforms from ICU beds to databases and a web-based user interface. Machine learning algorithms process the data streams and send outputs to the user interface. The architecture and capabilities of the platform are described. Since 2016, the platform has processed over 89 billion data points (N = 979 patients) from 200 signals (0.5-500 Hz) and laboratory analyses (once a day). We present an infrastructure-based framework for deploying and validating algorithms for critical care. The ICU Cockpit is a Big Data platform for critical care medicine, especially for multimodal waveform data. Uniquely, it allows algorithms to seamlessly integrate into the live data stream to produce clinical decision support and predictions in clinical practice.

Machine Learning and Intracranial Aneurysms: From Detection to Outcome Prediction

Machine learning (ML) is a rapidly rising research tool in biomedical sciences whose applications include segmentation, classification, disease detection, and outcome prediction. With respect to traditional statistical methods, ML algorithms have the potential to learn and improve their predictive performance when fed with large data sets without the need of being specifically programmed. In recent years, this technology has been increasingly applied for tackling clinical issues in intracranial aneurysm (IA) research. Several studies attempted to provide reliable models for enhanced aneurysm detection. Convolutional neural networks trained with variable degrees of human interaction on data from diverse imaging modalities showed high sensitivity in aneurysm detection tasks, also outperforming expert image analysis. Algorithms were also shown to differentiate ruptured from unruptured IAs, with however limited clinical relevance. For prediction of rupture and stability assessment, ML was preliminarily shown to achieve better performance compared to conventional statistical methods and existing risk scores. ML-based complication and functional outcome prediction in the event of SAH have been more extensively reported, in contrast with periprocedural outcome investigation in unruptured IA patients. ML has the potential to be a game changer in IA patient management. Currently clinical translation of experimental results is limited.

Impact of subcallosal artery origin and A1 asymmetry on surgical outcomes of anterior communicating artery aneurysms

BACKGROUND

Surgical clipping of anterior communicating artery (ACoA) aneurysms remains challenging due to their complex anatomy. Anatomical risk factors for ACoA aneurysm surgery require further elucidation. The aim of this study is to investigate whether proximity of the midline perforating artery, subcallosal artery (SubCA), and associated anomaly of the ACoA complex affect functional outcomes of ACoA aneurysm surgery.

METHODS

A total of 92 patients with both unruptured and ruptured ACoA aneurysms, who underwent surgical clipping, were retrospectively analyzed from a multicenter, observational cohort database. Association of ACoA anatomy with SubCA origin at the aneurysmal neck under microsurgical observation was analyzed in the interhemispheric approach subgroup (n = 56). Then, we evaluated whether anatomical factors associated with SubCA neck origin affected surgical outcomes in the entire cohort (both interhemispheric and pterional approaches, n = 92).

RESULTS

In the interhemispheric approach cohort, combination of A1 asymmetry and aneurysmal size ≥ 5.0 mm was stratified to have the highest probability of the SubCA neck origin by a decision tree analysis. Then, among the entire cohort using either interhemispheric or pterional approach, combination of A1 asymmetry and aneurysmal size ≥ 5.0 mm was significantly associated with poor functional outcomes by multivariable logistic regression analysis (OR 6.76; 95% CI 1.19-38.5; p = 0.03) as compared with A1 symmetry group in the acute subarachnoid hemorrhage settings.

CONCLUSION

Combination of A1 asymmetry and larger aneurysmal size was significantly associated with SubCA aneurysmal neck origin and poor functional outcomes in ACoA aneurysm surgery. Interhemispheric approach may be proposed to provide a wider and unobstructed view of SubCA for ACoA aneurysms with this high-risk anatomical variant.

CRP (C-Reactive Protein) in Outcome Prediction After Subarachnoid Hemorrhage and the Role of Machine Learning

BACKGROUND AND PURPOSE

Outcome prediction after aneurysmal subarachnoid hemorrhage (aSAH) is challenging. CRP (C-reactive protein) has been reported to be associated with outcome, but it is unclear if this is independent of other predictors and applies to aSAH of all grades. Therefore, the role of CRP in aSAH outcome prediction models is unknown. The purpose of this study is to assess if CRP is an independent predictor of outcome after aSAH, develop new prognostic models incorporating CRP, and test whether these can be improved by application of machine learning.

METHODS

This was an individual patient-level analysis of data from patients within 72 hours of aSAH from 2 prior studies. A panel of statistical learning methods including logistic regression, random forest, and support vector machines were used to assess the relationship between CRP and modified Rankin Scale. Models were compared with the full Subarachnoid Hemmorhage International Trialists' (SAHIT) prediction tool of outcome after aSAH and internally validated using cross-validation.

RESULTS

One thousand and seventeen patients were included for analysis. CRP on the first day after ictus was an independent predictor of outcome. The full SAHIT model achieved an area under the receiver operator characteristics curve (AUC) of 0.831. Addition of CRP to the predictors of the full SAHIT model improved model performance (AUC, 0.846, P=0.01). This improvement was not enhanced when learning was performed using a random forest (AUC, 0.807), but was with a support vector machine (AUC of 0.960, P <0.001).

CONCLUSIONS

CRP is an independent predictor of outcome after aSAH. Its inclusion in prognostic models improves performance, although the magnitude of improvement is probably insufficient to be relevant clinically on an individual patient level, and of more relevance in research. Greater improvements in model performance are seen with support vector machines but these models have the highest classification error rate on internal validation and require external validation and calibration.

Interleukin 6 and Aneurysmal Subarachnoid Hemorrhage. A Narrative Review

Interleukin 6 (IL-6) is a prominent proinflammatory cytokine. Neuroinflammation in general, and IL-6 signaling in particular, appear to play a major role in the pathobiology and pathophysiology of aneurysm formation and aneurysmal subarachnoid hemorrhage (SAH). Most importantly, elevated IL-6 CSF (rather than serum) levels appear to correlate with delayed cerebral ischemia (DCI, “vasospasm”) and secondary (“vasospastic”) infarctions. IL-6 CSF levels may also reflect other forms of injury to the brain following SAH, i.e., early brain damage and septic complications of SAH and aneurysm treatment. This would explain why many researchers have found an association between IL-6 levels and patient outcomes. These findings clearly suggest CSF IL-6 as a candidate biomarker in SAH patients. However, at this point, discrepant findings in variable study settings, as well as timing and other issues, e.g., defining proper clinical endpoints (i.e., secondary clinical deterioration vs. angiographic vasospasm vs. secondary vasospastic infarct) do not allow for its routine use. It is also tempting to speculate about potential therapeutic measures targeting elevated IL-6 CSF levels and neuroinflammation in SAH patients. Corticosteroids and anti-platelet drugs are indeed used in many SAH cases (not necessarily with the intention to interfere with detrimental inflammatory signaling), however, no convincing benefit has been demonstrated yet. The lack of a robust clinical perspective against the background of a relatively large body of data linking IL-6 and neuroinflammation with the pathophysiology of SAH is somewhat disappointing. One underlying reason might be that most relevant studies only report correlative data. The specific molecular pathways behind elevated IL-6 levels in SAH patients and their various interactions still remain to be delineated. We are optimistic that future research in this field will result in a better understanding of the role of neuroinflammation in the pathophysiology of SAH, which in turn, will translate into the identification of suitable biomarkers and even potential therapeutic targets.

Development of a Complication- and Treatment-Aware Prediction Model for Favorable Functional Outcome in Aneurysmal Subarachnoid Hemorrhage Based on Machine Learning

BACKGROUND

Current prognostic tools in aneurysmal subarachnoid hemorrhage (aSAH) are constrained by being primarily based on patient and disease characteristics on admission.

OBJECTIVE

To develop and validate a complication- and treatment-aware outcome prediction tool in aSAH.

METHODS

This cohort study included data from an ongoing prospective nationwide multicenter registry on all aSAH patients in Switzerland (Swiss SOS [Swiss Study on aSAH]; 2009-2015). We trained supervised machine learning algorithms to predict a binary outcome at discharge (modified Rankin scale [mRS] ≤ 3: favorable; mRS 4-6: unfavorable). Clinical and radiological variables on admission ("Early" Model) as well as additional variables regarding secondary complications and disease management ("Late" Model) were used. Performance of both models was assessed by classification performance metrics on an out-of-sample test dataset.

RESULTS

Favorable functional outcome at discharge was observed in 1156 (62.0%) of 1866 patients. Both models scored a high accuracy of 75% to 76% on the test set. The "Late" outcome model outperformed the "Early" model with an area under the receiver operator characteristics curve (AUC) of 0.85 vs 0.79, corresponding to a specificity of 0.81 vs 0.70 and a sensitivity of 0.71 vs 0.79, respectively.

CONCLUSION

Both machine learning models show good discrimination and calibration confirmed on application to an internal test dataset of patients with a wide range of disease severity treated in different institutions within a nationwide registry. Our study indicates that the inclusion of variables reflecting the clinical course of the patient may lead to outcome predictions with superior predictive power compared to a model based on admission data only.

Massive transfusion prediction in patients with multiple trauma by decision tree: a retrospective analysis

Early initial massive transfusion protocol and blood transfusion can reduce patient mortality, however accurately identifying the risk of massive transfusion (MT) remains a major challenge in severe trauma patient therapy. We retrospectively analyzed clinical data of severe trauma patients with and without MT. Based on analysis results, we established a MT prediction model of clinical and laboratory data by using the decision tree algorithm in patients with multiple trauma. Our results demonstrate that shock index, injury severity score, international normalized ratio, and pelvis fracture were the most significant risk factors of MT. These four indexes were incorporated into the prediction model, and the model was validated by using the testing dataset. Moreover, the sensitivity, specificity, accuracy and area under curve values of prediction model for MT risk prediction were 60%, 92%, 90% and 0.85. Our study provides an easy and understandable classification rules for identifying risk factors associated with MT that may be useful for promoting trauma management.

Predicting Long-Term Outcomes After Poor-Grade Aneurysmal Subarachnoid Hemorrhage Using Decision Tree Modeling

BACKGROUND

Despite advances in the treatment of poor-grade aneurysmal subarachnoid hemorrhage (aSAH), predicting the long-term outcome of aSAH remains challenging, although essential.

OBJECTIVE

To predict long-term outcomes after poor-grade aSAH using decision tree modeling.

METHODS

This was a retrospective analysis of a prospective multicenter observational registry of patients with poor-grade aSAH with a World Federation of Neurosurgical Societies (WFNS) grade IV or V. Outcome was assessed by the modified Rankin Scale (mRS) at 12 mo, and an unfavorable outcome was defined as an mRS of 4 or 5 or death. Long-term prognostic models were developed using multivariate logistic regression and decision tree algorithms. An additional independent testing dataset was collected for external validation. Overall accuracy, sensitivity, specificity, and area under receiver operating characteristic curves (AUC) were used to assess model performance.

RESULTS

Of the 266 patients, 139 (52.3%) had an unfavorable outcome. Older age, absence of pupillary reactivity, lower Glasgow coma score (GCS), and higher modified Fisher grade were independent predictors of unfavorable outcome. Modified Fisher grade, pupillary reactivity, GCS, and age were used in the decision tree model, which achieved an overall accuracy of 0.833, sensitivity of 0.821, specificity of 0.846, and AUC of 0.88 in the internal test. There was similar predictive performance between the logistic regression and decision tree models. Both models achieved a high overall accuracy of 0.895 in the external test.

CONCLUSION

Decision tree model is a simple tool for predicting long-term outcomes after poor-grade aSAH and may be considered for treatment decision-making.

Development of machine learning models to prognosticate chronic shunt-dependent hydrocephalus after aneurysmal subarachnoid hemorrhage

BACKGROUND

Shunt-dependent hydrocephalus significantly complicates subarachnoid hemorrhage (SAH), and reliable prognosis methods have been sought in recent years to reduce morbidity and costs associated with delayed treatment or neglected onset. Machine learning (ML) defines modern data analysis techniques allowing accurate subject-based risk stratifications. We aimed at developing and testing different ML models to predict shunt-dependent hydrocephalus after aneurysmal SAH.

METHODS

We consulted electronic records of patients with aneurysmal SAH treated at our institution between January 2013 and March 2019. We selected variables for the models according to the results of the previous works on this topic. We trained and tested four ML algorithms on three datasets: one containing binary variables, one considering variables associated with shunt-dependency after an explorative analysis, and one including all variables. For each model, we calculated AUROC, specificity, sensitivity, accuracy, PPV, and also, on the validation set, the NPV and the Matthews correlation coefficient (ϕ).

RESULTS

Three hundred eighty-six patients were included. Fifty patients (12.9%) developed shunt-dependency after a mean follow-up of 19.7 (± 12.6) months. Complete information was retrieved for 32 variables, used to train the models. The best models were selected based on the performances on the validation set and were achieved with a distributed random forest model considering 21 variables, with a ϕ = 0.59, AUC = 0.88; sensitivity and specificity of 0.73 (C.I.: 0.39-0.94) and 0.92 (C.I.: 0.84-0.97), respectively; PPV = 0.59 (0.38-0.77); and NPV = 0.96 (0.90-0.98). Accuracy was 0.90 (0.82-0.95).

CONCLUSIONS

Machine learning prognostic models allow accurate predictions with a large number of variables and a more subject-oriented prognosis. We identified a single best distributed random forest model, with an excellent prognostic capacity (ϕ = 0.58), which could be especially helpful in identifying low-risk patients for shunt-dependency.

Prediction of Clinical Outcome at Discharge After Rupture of Anterior Communicating Artery Aneurysm Using the Random Forest Technique

Background: Aneurysmal subarachnoid hemorrhage (SAH) is a devastating disease. Anterior communicating artery (ACoA) aneurysm is the most frequent location of intracranial aneurysms. The purpose of this study is to predict the clinical outcome at discharge after rupture of ACoA aneurysms using the random forest machine learning technique.

Methods: A total of 607 patients with ruptured ACoA aneurysms were included in this study between December 2007 and January 2016. In addition to basic clinical variables, 12 aneurysm morphologic parameters were evaluated. A multivariate logistic regression analysis was performed to determine the independent predictors of poor outcome. Of the 607 patients, 485 patients were randomly selected for training and the remaining for internal testing. The random forest model was developed using the training data set. An additional 202 patients from February 2016 to December 2017 were collected for externally validating the model. The prediction performance of the random forest model was compared with two radiologists.

Results: Patients' age (odds ratio [OR] = 1.04), ventilated breathing status (OR = 4.23), World Federation of Neurosurgical Societies (WFNS) grade (OR = 2.13), and Fisher grade (OR = 1.50) are significantly associated with poor outcome. None of the investigated morphological parameters of ACoA aneurysm is an independent predictor of poor outcome. The developed random forest model achieves sensitivities of 78.3% for internal test and 73.8% for external test. The areas under receiver operating characteristic (ROC) curve of the random forest model were 0.90 for the internal test and 0.84 for the external test. Both sensitivities and areas under ROC curves of our model are superior to those of two raters in both internal and external tests.

Conclusions: The random forest model presents good performance in predicting the outcome after rupture of ACoA aneurysms, which may aid in clinical decision making.

Feasible Study on Intracranial Hemorrhage Detection and Classification using a CNN-LSTM Network

Intracranial hemorrhage (ICH) is a life-threatening condition, the outcome of which is associated with stroke, trauma, aneurysm, vascular malformations, high blood pressure, illicit drugs and blood clotting disorders. In this study, we presented the feasibility of the automatic identification and classification of ICH using a head CT image based on deep learning technique. The subtypes of ICH for the classification was intraparenchymal, intraventricular, subarachnoid, subdural and epidural. We first performed windowing to provide three different images: brain window, bone window and subdural window, and trained 4,516,842 head CT images using CNN-LSTM model. We used the Xception model for the deep CNN, and 64 nodes and 32 timesteps for LSTM. For the performance evaluation, we tested 727,392 head CT images, and found the resultant weighted multi-label logarithmic loss was 0.07528. We believe that our proposed method enhances the accuracy of ICH identification and classification and can assist radiologists in the interpretation of head CT images, particularly for brain-related quantitative analysis.

Stability Assessment of Intracranial Aneurysms Using Machine Learning Based on Clinical and Morphological Features

Machine learning (ML) as a novel approach could help clinicians address the challenge of accurate stability assessment of unruptured intracranial aneurysms (IAs). We developed multiple ML models for IA stability assessment and compare their performances. We enrolled 1897 consecutive patients with unstable (n = 528) and stable (n = 1539) IAs. Thirteen patient-specific clinical features and eighteen aneurysm morphological features were extracted to generate support vector machine (SVM), random forest (RF), and feed-forward artificial neural network (ANN) models. The discriminatory performances of the models were compared with statistical logistic regression (LR) model and the PHASES score in IA stability assessment. Based on the receiver operating characteristic (ROC) curve and area under the curve (AUC) values for each model in the test set, the AUC values for RF, SVM, and ANN were 0.850 (95% CI 0.806-0.893), 0.858 (95 %CI 0.816-0.900), and 0.867 (95% CI 0.828-0.906), demonstrating good discriminatory ability. All ML models exhibited superior performance compared with the statistical LR and the PHASES score (the AUC values were 0.830 and 0.589, respectively; RF versus PHASES, P < 0.001; RF versus LR, P = 0.038). Important features contributing to the stability discrimination included three clinical features (location, sidewall/bifurcation type, and presence of symptoms) and three morphological features (undulation index, height-width ratio, and irregularity). These findings demonstrate the potential of ML to augment the clinical decision-making process for IA stability assessment, which may enable more optimal management for patients with IAs in the future.

Machine learning applications for the prediction of surgical site infection in neurological operations

OBJECTIVE

Surgical site infection (SSI) following a neurosurgical operation is a complication that impacts morbidity, mortality, and economics. Currently, machine learning (ML) algorithms are used for outcome prediction in various neurosurgical aspects. The implementation of ML algorithms to learn from medical data may help in obtaining prognostic information on diseases, especially SSIs. The purpose of this study was to compare the performance of various ML models for predicting surgical infection after neurosurgical operations.

METHODS

A retrospective cohort study was conducted on patients who had undergone neurosurgical operations at tertiary care hospitals between 2010 and 2017. Supervised ML algorithms, which included decision tree, naive Bayes with Laplace correction, k-nearest neighbors, and artificial neural networks, were trained and tested as binary classifiers (infection or no infection). To evaluate the ML models from the testing data set, their sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), as well as their accuracy, receiver operating characteristic curve, and area under the receiver operating characteristic curve (AUC) were analyzed.

RESULTS

Data were available for 1471 patients in the study period. The SSI rate was 4.6%, and the type of SSI was superficial, deep, and organ/space in 1.2%, 0.8%, and 2.6% of cases, respectively. Using the backward stepwise method, the authors determined that the significant predictors of SSI in the multivariable Cox regression analysis were postoperative CSF leakage/subgaleal collection (HR 4.24, p < 0.001) and postoperative fever (HR 1.67, p = 0.04). Compared with other ML algorithms, the naive Bayes had the highest performance with sensitivity at 63%, specificity at 87%, PPV at 29%, NPV at 96%, and AUC at 76%.

CONCLUSIONS

The naive Bayes algorithm is highlighted as an accurate ML method for predicting SSI after neurosurgical operations because of its reasonable accuracy. Thus, it can be used to effectively predict SSI in individual neurosurgical patients. Therefore, close monitoring and allocation of treatment strategies can be informed by ML predictions in general practice.

C-reactive protein to albumin ratio is a key indicator in a predictive model for anastomosis leakage after esophagectomy: Application of classification and regression tree analysis

Background

Anastomotic leakage (AL), a serious complication after esophagectomy, might impair patient quality of life, prolong hospital stay, and even lead to surgery‐related death. The aim of this study was to show a novel decision model based on classification and regression tree (CART) analysis for the prediction of postoperative AL among patients who have undergone esophagectomy.

Methods

A total of 450 patients (training set: 356; test set: 94) with perioperative information were included. A decision tree model was established to identify the predictors of AL in the training set, which was validated in the test set. A receiver operating characteristic curve was also created to illustrate the diagnostic ability of the decision model.

Results

A total of 12.2% (55/450) of the 450 patients suffered AL, which was diagnosed at median postoperative day 7 (range: 6–16). The decision tree model, containing surgical duration, postoperative lymphocyte count, and postoperative C‐reactive protein to albumin ratio, was established by CART analysis. Among the three variables, the postoperative C‐reactive protein to albumin ratio was identified as the most important indicator in the CART model with normalized importance of 100%. According to the results validated in the test set, the sensitivity, specificity, positive and negative predictive value, and diagnostic accuracy of the prediction model were 80%, 98.8%, 88.9%, 97.6%, and 96.8%, respectively. Moreover, the area under the receiver operating characteristic curve was 0.95.

Conclusion

The decision model based on CART analysis presented good performance for predicting AL, and might allow the early identification of patients at high risk.