Combination of Radiological and Clinical Baseline Data for Outcome Prediction of Patients With an Acute Ischemic Stroke.

Background: Accurate prediction of clinical outcome is of utmost importance for choices regarding the endovascular treatment (EVT) of acute stroke. Recent studies on the prediction modeling for stroke focused mostly on clinical characteristics and radiological scores available at baseline. Radiological images are composed of millions of voxels, and a lot of information can be lost when representing this information by a single value. Therefore, in this study we aimed at developing prediction models that take into account the whole imaging data combined with clinical data available at baseline. Methods: We included 3,279 patients from the MR CLEAN Registry; a prospective, observational, multicenter registry of patients with ischemic stroke treated with EVT. We developed two approaches to combine the imaging data with the clinical data. The first approach was based on radiomics features, extracted from 70 atlas regions combined with the clinical data to train machine learning models. For the second approach, we trained 3D deep learning models using the whole images and the clinical data. Models trained with the clinical data only were compared with models trained with the combination of clinical and image data. Finally, we explored feature importance plots for the best models and identified many known variables and image features/brain regions that were relevant in the model decision process. Results: From 3,279 patients included, 1,241 (37%) patients had a good functional outcome [modified Rankin Scale (mRS) ≤ 2] and 1,954 (60%) patients had good reperfusion [modified Thrombolysis in Cerebral Infarction (eTICI) ≥ 2b]. There was no significant improvement by combining the image data to the clinical data for mRS prediction [mean area under the receiver operating characteristic (ROC) curve (AUC) of 0.81 vs. 0.80] above using the clinical data only, regardless of the approach used. Regarding predicting reperfusion, there was a significant improvement when image and clinical features were combined (mean AUC of 0.54 vs. 0.61), with the highest AUC obtained by the deep learning approach. Conclusions: The combination of radiomics and deep learning image features with clinical data significantly improved the prediction of good reperfusion. The visualization of prediction feature importance showed both known and novel clinical and imaging features with predictive values.

Prediction of Atrial Fibrillation Recurrence after Thoracoscopic Surgical Ablation Using Machine Learning Techniques.

Thoracoscopic surgical ablation (SA) for atrial fibrillation (AF) has shown to be an effective treatment to restore sinus rhythm in patients with advanced AF. Identifying patients who will not benefit from this procedure would be valuable to improve personalized AF therapy. Machine learning (ML) techniques may assist in the improvement of clinical prediction models for patient selection. The aim of this study is to investigate how available baseline characteristics predict AF recurrence after SA using ML techniques. One-hundred-sixty clinical baseline variables were collected from 446 AF patients undergoing SA in our tertiary referral center. Multiple ML models were trained on five outcome measurements, including either all or a number of key variables selected by using the least absolute shrinkage and selection operator (LASSO). There was no difference in model performance between different ML techniques or outcome measurements. Variable selection significantly improved model performance (AUC: 0.73, 95% CI: 0.68-0.77). Subgroup analysis showed a higher model performance in younger patients (<55 years, AUC: 0.82 vs. >55 years, AUC 0.66). Recurrences of AF after SA can be predicted best when using a selection of baseline characteristics, particularly in young patients.

Predicting Success of a Digital Self-Help Intervention for Alcohol and Substance Use With Machine Learning.

BACKGROUND: Digital self-help interventions for reducing the use of alcohol tobacco and other drugs (ATOD) have generally shown positive but small effects in controlling substance use and improving the quality of life of participants. Nonetheless, low adherence rates remain a major drawback of these digital interventions, with mixed results in (prolonged) participation and outcome. To prevent non-adherence, we developed models to predict success in the early stages of an ATOD digital self-help intervention and explore the predictors associated with participant's goal achievement. METHODS: We included previous and current participants from a widely used, evidence-based ATOD intervention from the Netherlands (Jellinek Digital Self-help). Participants were considered successful if they completed all intervention modules and reached their substance use goals (i.e., stop/reduce). Early dropout was defined as finishing only the first module. During model development, participants were split per substance (alcohol, tobacco, cannabis) and features were computed based on the log data of the first 3 days of intervention participation. Machine learning models were trained, validated and tested using a nested k-fold cross-validation strategy. RESULTS: From the 32,398 participants enrolled in the study, 80% of participants did not complete the first module of the intervention and were excluded from further analysis. From the remaining participants, the percentage of success for each substance was 30% for alcohol, 22% for cannabis and 24% for tobacco. The area under the Receiver Operating Characteristic curve was the highest for the Random Forest model trained on data from the alcohol and tobacco programs (0.71 95%CI 0.69-0.73) and (0.71 95%CI 0.67-0.76), respectively, followed by cannabis (0.67 95%CI 0.59-0.75). Quitting substance use instead of moderation as an intervention goal, initial daily consumption, no substance use on the weekends as a target goal and intervention engagement were strong predictors of success. DISCUSSION: Using log data from the first 3 days of intervention use, machine learning models showed positive results in identifying successful participants. Our results suggest the models were especially able to identify participants at risk of early dropout. Multiple variables were found to have high predictive value, which can be used to further improve the intervention.

Predicting mortality of individual patients with COVID-19: a multicentre Dutch cohort.

OBJECTIVE: Develop and validate models that predict mortality of patients diagnosed with COVID-19 admitted to the hospital. DESIGN: Retrospective cohort study. SETTING: A multicentre cohort across 10 Dutch hospitals including patients from 27 February to 8 June 2020. PARTICIPANTS: SARS-CoV-2 positive patients (age ≥18) admitted to the hospital. MAIN OUTCOME MEASURES: 21-day all-cause mortality evaluated by the area under the receiver operator curve (AUC), sensitivity, specificity, positive predictive value and negative predictive value. The predictive value of age was explored by comparison with age-based rules used in practice and by excluding age from the analysis. RESULTS: 2273 patients were included, of whom 516 had died or discharged to palliative care within 21 days after admission. Five feature sets, including premorbid, clinical presentation and laboratory and radiology values, were derived from 80 features. Additionally, an Analysis of Variance (ANOVA)-based data-driven feature selection selected the 10 features with the highest F values: age, number of home medications, urea nitrogen, lactate dehydrogenase, albumin, oxygen saturation (%), oxygen saturation is measured on room air, oxygen saturation is measured on oxygen therapy, blood gas pH and history of chronic cardiac disease. A linear logistic regression and non-linear tree-based gradient boosting algorithm fitted the data with an AUC of 0.81 (95% CI 0.77 to 0.85) and 0.82 (0.79 to 0.85), respectively, using the 10 selected features. Both models outperformed age-based decision rules used in practice (AUC of 0.69, 0.65 to 0.74 for age >70). Furthermore, performance remained stable when excluding age as predictor (AUC of 0.78, 0.75 to 0.81). CONCLUSION: Both models showed good performance and had better test characteristics than age-based decision rules, using 10 admission features readily available in Dutch hospitals. The models hold promise to aid decision-making during a hospital bed shortage.

Improving electrocardiogram-based detection of rare genetic heart disease using transfer learning: An application to phospholamban p.Arg14del mutation carriers.

The pathogenic mutation p.Arg14del in the gene encoding Phospholamban (PLN) is known to cause cardiomyopathy and leads to increased risk of sudden cardiac death. Automatic tools might improve the detection of patients with this rare disease. Deep learning is currently the state-of-the-art in signal processing but requires large amounts of data to train the algorithms. In situations with relatively small amounts of data, like PLN, transfer learning may improve accuracy. We propose an ECG-based detection of the PLN mutation using transfer learning from a model originally trained for sex identification. The sex identification model was trained with 256,278 ECGs and subsequently finetuned for PLN detection (155 ECGs of patients with PLN) with two control groups: a balanced age/sex matched group and a randomly selected imbalanced population. The data was split in 10 folds and 20% of the training data was used for validation and early stopping. The models were evaluated with the area under the receiver operating characteristic curve (AUROC) of the testing data. We used gradient activation for explanation of the prediction models. The models trained with transfer learning outperformed the models trained from scratch for both the balanced (AUROC 0.87 vs AUROC 0.71) and imbalanced (AUROC 0.0.90 vs AUROC 0.65) population. The proposed approach was able to improve the accuracy of a rare disease detection model by transfer learning information from a non-manual annotated and abundant label with only limited data available.

Computer versus cardiologist: Is a machine learning algorithm able to outperform an expert in diagnosing a phospholamban p.Arg14del mutation on the electrocardiogram?

BACKGROUND: Phospholamban (PLN) p.Arg14del mutation carriers are known to develop dilated and/or arrhythmogenic cardiomyopathy, and typical electrocardiographic (ECG) features have been identified for diagnosis. Machine learning is a powerful tool used in ECG analysis and has shown to outperform cardiologists. OBJECTIVES: We aimed to develop machine learning and deep learning models to diagnose PLN p.Arg14del cardiomyopathy using ECGs and evaluate their accuracy compared to an expert cardiologist. METHODS: We included 155 adult PLN mutation carriers and 155 age- and sex-matched control subjects. Twenty-one PLN mutation carriers (13.4%) were classified as symptomatic (symptoms of heart failure or malignant ventricular arrhythmias). The data set was split into training and testing sets using 4-fold cross-validation. Multiple models were developed to discriminate between PLN mutation carriers and control subjects. For comparison, expert cardiologists classified the same data set. The best performing models were validated using an external PLN p.Arg14del mutation carrier data set from Murcia, Spain (n = 50). We applied occlusion maps to visualize the most contributing ECG regions. RESULTS: In terms of specificity, expert cardiologists (0.99) outperformed all models (range 0.53-0.81). In terms of accuracy and sensitivity, experts (0.28 and 0.64) were outperformed by all models (sensitivity range 0.65-0.81). T-wave morphology was most important for classification of PLN p.Arg14del carriers. External validation showed comparable results, with the best model outperforming experts. CONCLUSION: This study shows that machine learning can outperform experienced cardiologists in the diagnosis of PLN p.Arg14del cardiomyopathy and suggests that the shape of the T wave is of added importance to this diagnosis.

qTICI: Quantitative assessment of brain tissue reperfusion on digital subtraction angiograms of acute ischemic stroke patients.

BACKGROUND: The Thrombolysis in Cerebral Infarction (TICI) scale is an important outcome measure to evaluate the quality of endovascular stroke therapy. The TICI scale is ordinal and observer-dependent, which may result in suboptimal prediction of patient outcome and inconsistent reperfusion grading. AIMS: We present a semi-automated quantitative reperfusion measure (quantified TICI (qTICI)) using image processing techniques based on the TICI methodology. METHODS: We included patients with an intracranial proximal large vessel occlusion with complete, good quality runs of anteroposterior and lateral digital subtraction angiography from the MR CLEAN Registry. For each vessel occlusion, we identified the target downstream territory and automatically segmented the reperfused area in the target downstream territory on final digital subtraction angiography. qTICI was defined as the percentage of reperfused area in target downstream territory. The value of qTICI and extended TICI (eTICI) in predicting favorable functional outcome (modified Rankin Scale 0-2) was compared using area under receiver operating characteristics curve and binary logistic regression analysis unadjusted and adjusted for known prognostic factors. RESULTS: In total, 408 patients with M1 or internal carotid artery occlusion were included. The median qTICI was 78 (interquartile range 58-88) and 215 patients (53%) had an eTICI of 2C or higher. qTICI was comparable to eTICI in predicting favorable outcome with area under receiver operating characteristics curve of 0.63 vs. 0.62 (P = 0.8) and 0.87 vs. 0.86 (P = 0.87), for the unadjusted and adjusted analysis, respectively. In the adjusted regression analyses, both qTICI and eTICI were independently associated with functional outcome. CONCLUSION: qTICI provides a quantitative measure of reperfusion with similar prognostic value for functional outcome to eTICI score.

Predicting Poor Outcome Before Endovascular Treatment in Patients With Acute Ischemic Stroke.

Background: Although endovascular treatment (EVT) has greatly improved outcomes in acute ischemic stroke, still one third of patients die or remain severely disabled after stroke. If we could select patients with poor clinical outcome despite EVT, we could prevent futile treatment, avoid treatment complications, and further improve stroke care. We aimed to determine the accuracy of poor functional outcome prediction, defined as 90-day modified Rankin Scale (mRS) score ≥5, despite EVT treatment. Methods: We included 1,526 patients from the MR CLEAN Registry, a prospective, observational, multicenter registry of ischemic stroke patients treated with EVT. We developed machine learning prediction models using all variables available at baseline before treatment. We optimized the models for both maximizing the area under the curve (AUC), reducing the number of false positives. Results: From 1,526 patients included, 480 (31%) of patients showed poor outcome. The highest AUC was 0.81 for random forest. The highest area under the precision recall curve was 0.69 for the support vector machine. The highest achieved specificity was 95% with a sensitivity of 34% for neural networks, indicating that all models contained false positives in their predictions. From 921 mRS 0-4 patients, 27-61 (3-6%) were incorrectly classified as poor outcome. From 480 poor outcome patients in the registry, 99-163 (21-34%) were correctly identified by the models. Conclusions: All prediction models showed a high AUC. The best-performing models correctly identified 34% of the poor outcome patients at a cost of misclassifying 4% of non-poor outcome patients. Further studies are necessary to determine whether these accuracies are reproducible before implementation in clinical practice.

A morphology based deep learning model for atrial fibrillation detection using single cycle electrocardiographic samples.

BACKGROUND: Deep learning (DL) has shown promising results in improving atrial fibrillation (AF) detection algorithms. However, these models are often criticized because of their "black box" nature. AIM: To develop a morphology based DL model to discriminate AF from sinus rhythm (SR), and to visualize which parts of the ECG are used by the model to derive to the right classification. METHODS: We pre-processed raw data of 1469 ECGs in AF or SR, of patients with a history AF. Input data was generated by normalizing all single cycles (SC) of one ECG lead to SC-ECG samples by 1) centralizing the R wave or 2) scaling from R-to- R wave. Different DL models were trained by splitting the data in a training, validation and test set. By using a DL based heat mapping technique we visualized those areas of the ECG used by the classifier to come to the correct classification. RESULTS: The DL model with the best performance was a feedforward neural network trained by SC-ECG samples on a R-to-R wave basis of lead II, resulting in an accuracy of 0.96 and F1-score of 0.94. The onset of the QRS complex proved to be the most relevant area for the model to discriminate AF from SR. CONCLUSION: The morphology based DL model developed in this study was able to discriminate AF from SR with a very high accuracy. DL model visualization may help clinicians gain insights into which (unrecognized) ECG features are most sensitive to discriminate AF from SR.

Predicting Delayed Cerebral Ischemia with Quantified Aneurysmal Subarachnoid Blood Volume.

BACKGROUND: The amount of blood detected on brain computed tomography scan is frequently used in prediction models for delayed cerebral ischemia (DCI) in patients with aneurysmal subarachnoid hemorrhage (aSAH). These models, which include coarse grading scales to assess the amount of blood, have only moderate predictive value. Therefore, we aimed to develop a predictive model for DCI including automatically quantified total blood volume (TBV). METHODS: We included patients from a prospective aSAH registry. TBV was assessed with an automatic hemorrhage quantification algorithm. The outcome measure was clinical deterioration due to DCI. Clinical and radiologic variables were included in a logistic regression model. The final model was selected by bootstrapped backward selection and internally validated by assessing the optimism-corrected R2 value, c-statistic, and calibration plot. The c-statistic of the TBV model was compared with models that used the (modified) Fisher scale instead. RESULTS: We included 369 patients. After backward selection, only TBV was included in the final model. The internally validated R2 value was 6%, and the c-statistic was 0.64. The c-statistic of the TBV model was higher than both the Fisher scale model (0.56; P < 0.001) and the modified Fisher scale model (0.58; P < 0.05). CONCLUSIONS: In our registry, only TBV independently predicted DCI. TBV discriminated better than the (modified) Fisher scale, but still had only moderate value for predicting DCI. Our findings suggest that other factors need to be identified to achieve better accuracy for predicting DCI.

Predicting Outcome of Endovascular Treatment for Acute Ischemic Stroke: Potential Value of Machine Learning Algorithms.

Background: Endovascular treatment (EVT) is effective for stroke patients with a large vessel occlusion (LVO) of the anterior circulation. To further improve personalized stroke care, it is essential to accurately predict outcome after EVT. Machine learning might outperform classical prediction methods as it is capable of addressing complex interactions and non-linear relations between variables. Methods: We included patients from the Multicenter Randomized Clinical Trial of Endovascular Treatment for Acute Ischemic Stroke in the Netherlands (MR CLEAN) Registry, an observational cohort of LVO patients treated with EVT. We applied the following machine learning algorithms: Random Forests, Support Vector Machine, Neural Network, and Super Learner and compared their predictive value with classic logistic regression models using various variable selection methodologies. Outcome variables were good reperfusion (post-mTICI ≥ 2b) and functional independence (modified Rankin Scale ≤2) at 3 months using (1) only baseline variables and (2) baseline and treatment variables. Area under the ROC-curves (AUC) and difference of mean AUC between the models were assessed. Results: We included 1,383 EVT patients, with good reperfusion in 531 (38%) and functional independence in 525 (38%) patients. Machine learning and logistic regression models all performed poorly in predicting good reperfusion (range mean AUC: 0.53-0.57), and moderately in predicting 3-months functional independence (range mean AUC: 0.77-0.79) using only baseline variables. All models performed well in predicting 3-months functional independence using both baseline and treatment variables (range mean AUC: 0.88-0.91) with a negligible difference of mean AUC (0.01; 95%CI: 0.00-0.01) between best performing machine learning algorithm (Random Forests) and best performing logistic regression model (based on prior knowledge). Conclusion: In patients with LVO machine learning algorithms did not outperform logistic regression models in predicting reperfusion and 3-months functional independence after endovascular treatment. For all models at time of admission radiological outcome was more difficult to predict than clinical outcome.