Perfusion parameter map generation from TOF-MRA in stroke using generative adversarial networks.

PURPOSE: To generate perfusion parameter maps from Time-of-flight magnetic resonance angiography (TOF-MRA) images using artificial intelligence to provide an alternative to traditional perfusion imaging techniques. MATERIALS AND METHODS: This retrospective study included a total of 272 patients with cerebrovascular diseases; 200 with acute stroke (from 2010 to 2018), and 72 with steno-occlusive disease (from 2011 to 2014). For each patient the TOF MRA image and the corresponding Dynamic susceptibility contrast magnetic resonance imaging (DSC-MRI) were retrieved from the datasets. The authors propose an adapted generative adversarial network (GAN) architecture, 3D pix2pix GAN, that generates common perfusion maps (CBF, CBV, MTT, TTP, Tmax) from TOF-MRA images. The performance was evaluated by the structural similarity index measure (SSIM). For a subset of 20 patients from the acute stroke dataset, the Dice coefficient was calculated to measure the overlap between the generated and real hypoperfused lesions with a time-to-maximum (Tmax) > 6 s. RESULTS: The GAN model exhibited high visual overlap and performance for all perfusion maps in both datasets: acute stroke (mean SSIM 0.88-0.92, mean PSNR 28.48-30.89, mean MAE 0.02-0.04 and mean NRMSE 0.14-0.37) and steno-occlusive disease patients (mean SSIM 0.83-0.98, mean PSNR 23.62-38.21, mean MAE 0.01-0.05 and mean NRMSE 0.03-0.15). For the overlap analysis for lesions with Tmax>6 s, the median Dice coefficient was 0.49. CONCLUSION: Our AI model can successfully generate perfusion parameter maps from TOF-MRA images, paving the way for a non-invasive alternative for assessing cerebral hemodynamics in cerebrovascular disease patients. This method could impact the stratification of patients with cerebrovascular diseases. Our results warrant more extensive refinement and validation of the method.

The Impact of Multi-Institution Datasets on the Generalizability of Machine Learning Prediction Models in the ICU.

OBJECTIVES: To evaluate the transferability of deep learning (DL) models for the early detection of adverse events to previously unseen hospitals. DESIGN: Retrospective observational cohort study utilizing harmonized intensive care data from four public datasets. SETTING: ICUs across Europe and the United States. PATIENTS: Adult patients admitted to the ICU for at least 6 hours who had good data quality. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: Using carefully harmonized data from a total of 334,812 ICU stays, we systematically assessed the transferability of DL models for three common adverse events: death, acute kidney injury (AKI), and sepsis. We tested whether using more than one data source and/or algorithmically optimizing for generalizability during training improves model performance at new hospitals. We found that models achieved high area under the receiver operating characteristic (AUROC) for mortality (0.838-0.869), AKI (0.823-0.866), and sepsis (0.749-0.824) at the training hospital. As expected, AUROC dropped when models were applied at other hospitals, sometimes by as much as -0.200. Using more than one dataset for training mitigated the performance drop, with multicenter models performing roughly on par with the best single-center model. Dedicated methods promoting generalizability did not noticeably improve performance in our experiments. CONCLUSIONS: Our results emphasize the importance of diverse training data for DL-based risk prediction. They suggest that as data from more hospitals become available for training, models may become increasingly generalizable. Even so, good performance at a new hospital still depended on the inclusion of compatible hospitals during training.

Enhancing the prediction for shunt-dependent hydrocephalus after aneurysmal subarachnoid hemorrhage using a machine learning approach.

Early and reliable prediction of shunt-dependent hydrocephalus (SDHC) after aneurysmal subarachnoid hemorrhage (aSAH) may decrease the duration of in-hospital stay and reduce the risk of catheter-associated meningitis. Machine learning (ML) may improve predictions of SDHC in comparison to traditional non-ML methods. ML models were trained for CHESS and SDASH and two combined individual feature sets with clinical, radiographic, and laboratory variables. Seven different algorithms were used including three types of generalized linear models (GLM) as well as a tree boosting (CatBoost) algorithm, a Naive Bayes (NB) classifier, and a multilayer perceptron (MLP) artificial neural net. The discrimination of the area under the curve (AUC) was classified (0.7 ≤ AUC < 0.8, acceptable; 0.8 ≤ AUC < 0.9, excellent; AUC ≥ 0.9, outstanding). Of the 292 patients included with aSAH, 28.8% (n = 84) developed SDHC. Non-ML-based prediction of SDHC produced an acceptable performance with AUC values of 0.77 (CHESS) and 0.78 (SDASH). Using combined feature sets with more complex variables included than those incorporated in the scores, the ML models NB and MLP reached excellent performances, with an AUC of 0.80, respectively. After adding the amount of CSF drained within the first 14 days as a late feature to ML-based prediction, excellent performances were reached in the MLP (AUC 0.81), NB (AUC 0.80), and tree boosting model (AUC 0.81). ML models may enable clinicians to reliably predict the risk of SDHC after aSAH based exclusively on admission data. Future ML models may help optimize the management of SDHC in aSAH by avoiding delays in clinical decision-making.

An evaluation of performance measures for arterial brain vessel segmentation.

BACKGROUND: Arterial brain vessel segmentation allows utilising clinically relevant information contained within the cerebral vascular tree. Currently, however, no standardised performance measure is available to evaluate the quality of cerebral vessel segmentations. Thus, we developed a performance measure selection framework based on manual visual scoring of simulated segmentation variations to find the most suitable measure for cerebral vessel segmentation. METHODS: To simulate segmentation variations, we manually created non-overlapping segmentation errors common in magnetic resonance angiography cerebral vessel segmentation. In 10 patients, we generated a set of approximately 300 simulated segmentation variations for each ground truth image. Each segmentation was visually scored based on a predefined scoring system and segmentations were ranked based on 22 performance measures common in the literature. The correlation of visual scores with performance measure rankings was calculated using the Spearman correlation coefficient. RESULTS: The distance-based performance measures balanced average Hausdorff distance (rank = 1) and average Hausdorff distance (rank = 2) provided the segmentation rankings with the highest average correlation with manual rankings. They were followed by overlap-based measures such as Dice coefficient (rank = 7), a standard performance measure in medical image segmentation. CONCLUSIONS: Average Hausdorff distance-based measures should be used as a standard performance measure in evaluating cerebral vessel segmentation quality. They can identify more relevant segmentation errors, especially in high-quality segmentations. Our findings have the potential to accelerate the validation and development of novel vessel segmentation approaches.

Outcome prediction in aneurysmal subarachnoid hemorrhage: a comparison of machine learning methods and established clinico-radiological scores.

Reliable prediction of outcomes of aneurysmal subarachnoid hemorrhage (aSAH) based on factors available at patient admission may support responsible allocation of resources as well as treatment decisions. Radiographic and clinical scoring systems may help clinicians estimate disease severity, but their predictive value is limited, especially in devising treatment strategies. In this study, we aimed to examine whether a machine learning (ML) approach using variables available on admission may improve outcome prediction in aSAH compared to established scoring systems. Combined clinical and radiographic features as well as standard scores (Hunt & Hess, WFNS, BNI, Fisher, and VASOGRADE) available on patient admission were analyzed using a consecutive single-center database of patients that presented with aSAH (n = 388). Different ML models (seven algorithms including three types of traditional generalized linear models, as well as a tree bosting algorithm, a support vector machine classifier (SVMC), a Naive Bayes (NB) classifier, and a multilayer perceptron (MLP) artificial neural net) were trained for single features, scores, and combined features with a random split into training and test sets (4:1 ratio), ten-fold cross-validation, and 50 shuffles. For combined features, feature importance was calculated. There was no difference in performance between traditional and other ML applications using traditional clinico-radiographic features. Also, no relevant difference was identified between a combined set of clinico-radiological features available on admission (highest AUC 0.78, tree boosting) and the best performing clinical score GCS (highest AUC 0.76, tree boosting). GCS and age were the most important variables for the feature combination. In this cohort of patients with aSAH, the performance of functional outcome prediction by machine learning techniques was comparable to traditional methods and established clinical scores. Future work is necessary to examine input variables other than traditional clinico-radiographic features and to evaluate whether a higher performance for outcome prediction in aSAH can be achieved.

The choice of embedding media affects image quality, tissue R2* , and susceptibility behaviors in post-mortem brain MR microscopy at 7.0T.

PURPOSE: The quality and precision of post-mortem MRI microscopy may vary depending on the embedding medium used. To investigate this, our study evaluated the impact of 5 widely used media on: (1) image quality, (2) contrast of high spatial resolution gradient-echo (T1 and T2* -weighted) MR images, (3) effective transverse relaxation rate (R2* ), and (4) quantitative susceptibility measurements (QSM) of post-mortem brain specimens. METHODS: Five formaldehyde-fixed brain slices were scanned using 7.0T MRI in: (1) formaldehyde solution (formalin), (2) phosphate-buffered saline (PBS), (3) deuterium oxide (D2 O), (4) perfluoropolyether (Galden), and (5) agarose gel. SNR and contrast-to-noise ratii (SNR/CNR) were calculated for cortex/white matter (WM) and basal ganglia/WM regions. In addition, median R2* and QSM values were extracted from caudate nucleus, putamen, globus pallidus, WM, and cortical regions. RESULTS: PBS, Galden, and agarose returned higher SNR/CNR compared to formalin and D2 O. Formalin fixation, and its use as embedding medium for scanning, increased tissue R2* . Imaging with agarose, D2 O, and Galden returned lower R2* values than PBS (and formalin). No major QSM offsets were observed, although spatial variance was increased (with respect to R2* behaviors) for formalin and agarose. CONCLUSIONS: Embedding media affect gradient-echo image quality, R2* , and QSM in differing ways. In this study, PBS embedding was identified as the most stable experimental setup, although by a small margin. Agarose and Galden were preferred to formalin or D2 O embedding. Formalin significantly increased R2* causing noisier data and increased QSM variance.

Walsh-ordered hadamard time-encoded pseudocontinuous ASL (WH pCASL).

PURPOSE: Walsh ordering of Hadamard encoding-matrices and an additional averaging strategy are proposed for Hadamard-encoded pseudocontinuous arterial spin labeling (H-pCASL). In contrast to conventional H-pCASL the proposed method generates more perfusion-weighted images which are accessible already during a running experiment and even from incomplete sets of encoded images. THEORY: Walsh-ordered Hadamard matrices consist of fully decodable Hadamard submatrices. At any time during a measurement these submatrices may yield perfusion-weighted images, even at runtime and with incomplete datasets. The usage of an additional so-called mirrored matrix for averaging leads to more decodable subboli. METHODS: Perfusion-weighted images (five healthy volunteers) are generated using both a Walsh-ordered and a corresponding mirrored Hadamard matrix. To test their correctness they are compared with equivalent images from conventional multi postlabeling-delay (PLD) pCASL-measurements. RESULTS: All predicted perfusion-weighted images could be generated and correlated very well with multi-PLD images. Already small subsets of encoded images, acquired early during the measurement, yielded perfusion-weighted images. The mirrored matrix generates more perfusion-weighted images without time penalty. CONCLUSION: Early access to perfusion-weighted images despite incomplete datasets allows dynamic adaptation of parameters and increases robustness against artifacts. Thus, the approach may be well suited for clinical applications, where pathologies demand rapid parameter adaptation and increase the chance of artifacts. Magn Reson Med 76:1814-1824, 2016. © 2015 International Society for Magnetic Resonance in Medicine.

Clinical-Radiological Parameters Improve the Prediction of the Thrombolysis Time Window by Both MRI Signal Intensities and DWI-FLAIR Mismatch.

BACKGROUND: With regard to acute stroke, patients with unknown time from stroke onset are not eligible for thrombolysis. Quantitative diffusion weighted imaging (DWI) and fluid attenuated inversion recovery (FLAIR) MRI relative signal intensity (rSI) biomarkers have been introduced to predict eligibility for thrombolysis, but have shown heterogeneous results in the past. In the present work, we investigated whether the inclusion of easily obtainable clinical-radiological parameters would improve the prediction of the thrombolysis time window by rSIs and compared their performance to the visual DWI-FLAIR mismatch. METHODS: In a retrospective study, patients from 2 centers with proven stroke with onset <12 h were included. The DWI lesion was segmented and overlaid on ADC and FLAIR images. rSI mean and SD, were calculated as follows: (mean ROI value/mean value of the unaffected hemisphere). Additionally, the visual DWI-FLAIR mismatch was evaluated. Prediction of the thrombolysis time window was evaluated by the area-under-the-curve (AUC) derived from receiver operating characteristic (ROC) curve analysis. Factors such as the association of age, National Institutes of Health Stroke Scale, MRI field strength, lesion size, vessel occlusion and Wahlund-Score with rSI were investigated and the models were adjusted and stratified accordingly. RESULTS: In 82 patients, the unadjusted rSI measures DWI-mean and -SD showed the highest AUCs (AUC 0.86-0.87). Adjustment for clinical-radiological covariates significantly improved the performance of FLAIR-mean (0.91) and DWI-SD (0.91). The best prediction results based on the AUC were found for the final stratified and adjusted models of DWI-SD (0.94) and FLAIR-mean (0.96) and a multivariable DWI-FLAIR model (0.95). The adjusted visual DWI-FLAIR mismatch did not perform in a significantly worse manner (0.89). ADC-rSIs showed fair performance in all models. CONCLUSIONS: Quantitative DWI and FLAIR MRI biomarkers as well as the visual DWI-FLAIR mismatch provide excellent prediction of eligibility for thrombolysis in acute stroke, when easily obtainable clinical-radiological parameters are included in the prediction models.

Long-term trajectories of community integration: identification, characterization, and prediction using inpatient rehabilitation variables.

BACKGROUND: Community integration (CI) is often regarded as the foundation of rehabilitation endeavors after stroke; nevertheless, few studies have investigated the relationship between inpatient rehabilitation (clinical and demographic) variables and long-term CI. OBJECTIVES: To identify novel classes of patients having similar temporal patterns in CI and relate them to baseline features. METHODS: Retrospective observational cohort study analyzing (n = 287) adult patients with stroke admitted to rehabilitation between 2003 and 2018, including baseline Functional Independence Measure (FIM) at discharge, follow-ups (m = 1264) of Community Integration Questionnaire (CIQ) between 2006 and 2022. Growth mixture models (GMMs) were fitted to identify CI trajectories, and baseline predictors were identified using multivariate logistic regression (reporting AUC) with 10-fold cross validation. RESULTS: Each patient was assessed at 2.7 (2.2-3.7), 4.4 (3.7-5.6), and 6.2 (5.4-7.4) years after injury, 66% had a fourth assessment at 7.9 (6.8-8.9) years. GMM identified three classes of trajectories.Lowest CI (n=105, 36.6%): The lowest mean total CIQ; highest proportion of dysphagia (47.6%) and aphasia (46.7%), oldest at injury, largest length of stay (LOS), largest time to admission, and lowest FIM.Highest CI (n=63, 21.9%): The highest mean total CIQ, youngest, shortest LOS, highest education (27% university) highest FIM, and Intermediate CI (n=119, 41.5%): Intermediate mean total CIQ and FIM scores. Age at injury OR: 0.89 (0.85-0.93), FIM OR: 1.04 (1.02-1.07), hypertension OR: 2.86 (1.25-6.87), LOS OR: 0.98 (0.97-0.99), and high education OR: 3.05 (1.22-7.65) predicted highest CI, and AUC was 0.84 (0.76-0.93). CONCLUSION: Novel clinical (e.g. hypertension) and demographic (e.g. education) variables characterized and predicted long-term CI trajectories.

Predicting models for arm impairment: External validation of the Scandinavian models and identification of new predictors in post-acute stroke settings.

BACKGROUND: Post-stroke arm impairment at rehabilitation admission as predictor of discharge arm impairment was consistently reported as extremely useful. Several models for acute prediction exist (e.g. the Scandinavian), though lacking external validation and larger time-window admission assessments. OBJECTIVES: (1) use the 33 Fugl-Meyer Assessment-Upper Extremity (FMA-UE) individual items to predict total FMA-UE score at discharge of patients with ischemic stroke admitted to rehabilitation within 90 days post-injury, (2) use eight individual items (seven from the Scandinavian study plus the top predictor item from objective 1) to predict mild impairment (FMA-UE≥48) at discharge and (3) adjust the top three models from objective 2 with known confounders. METHODS: This was an observational study including 287 patients (from eight settings) admitted to rehabilitation (2009-2020). We applied regression models to candidate predictors, reporting adjusted R2, odds ratios and ROC-AUC using 10-fold cross-validation. RESULTS: We achieved good predictive power for the eight item-level models (AUC: 0.70-0.82) and for the three adjusted models (AUC: 0.85-0.88). We identified finger mass flexion as new item-level top predictor (AUC:0.88) and time to admission (OR = 0.9(0.9;1.0)) as only common significant confounder. CONCLUSION: Scandinavian item-level predictors are valid in a different context, finger mass flexion outperformed known predictors, days-to-admission predict discharge mild arm impairment.

You Can't Have AI Both Ways: Balancing Health Data Privacy and Access Fairly.

Artificial intelligence (AI) in healthcare promises to make healthcare safer, more accurate, and more cost-effective. Public and private actors have been investing significant amounts of resources into the field. However, to benefit from data-intensive medicine, particularly from AI technologies, one must first and foremost have access to data. It has been previously argued that the conventionally used "consent or anonymize approach" undermines data-intensive medicine, and worse, may ultimately harm patients. Yet, this is still a dominant approach in European countries and framed as an either-or choice. In this paper, we contrast the different data governance approaches in the EU and their advantages and disadvantages in the context of healthcare AI. We detail the ethical trade-offs inherent to data-intensive medicine, particularly the balancing of data privacy and data access, and the subsequent prioritization between AI and other effective health interventions. If countries wish to allocate resources to AI, they also need to make corresponding efforts to improve (secure) data access. We conclude that it is unethical to invest significant amounts of public funds into AI development whilst at the same time limiting data access through strict privacy measures, as this constitutes a waste of public resources. The "AI revolution" in healthcare can only realise its full potential if a fair, inclusive engagement process spells out the values underlying (trans) national data governance policies and their impact on AI development, and priorities are set accordingly.

The impact of coronavirus disease 2019 on emotional and behavioral stress of informal family caregivers of individuals with stroke or traumatic brain injury at chronic phase living in a Mediterranean setting.

INTRODUCTION: Even in nonpandemic times, persons with disabilities experience emotional and behavioral disturbances which are distressing for them and for their close persons. We aimed at comparing the levels of stress in emotional and behavioral aspects, before and during coronavirus disease 2019 (COVID-19), as reported by informal family caregivers of individuals with chronic traumatic brain injury (TBI) or stroke living in the community, considering two different stratifications of the recipients of care (cause and injury severity). METHODS: We conducted a STROBE-compliant prospective observational study analyzing informal caregivers of individuals with stroke (IC-STROKE) or traumatic brain injury (IC-TBI). IC-STROKE and IC-TBI were assessed in-person before and during COVID-19 online, using the Head Injury Behavior Scale (HIBS). The HIBS comprises behavioral and emotional subtotals (10 items each) and a total-HIBS. Comparisons were performed using the McNemar's test, Wilcoxon signed-rank test or t-test. Recipients of care were stratified according to their injury severity using the National Institutes of Health Stroke Scale (NIHSS) and the Glasgow Coma Scale (GCS). RESULTS: One hundred twenty-two informal caregivers (62.3% IC-STROKE and 37.7% IC-TBI) were assessed online between June 2020 and April 2021 and compared to their own assessments performed in-person 1.74 ± 0.88 years before the COVID-19 lockdown. IC-STROKE significantly increased their level of stress during COVID-19 in five emotional items (impatience, frequent complaining, often disputes topics, mood change and overly sensitive) and in one behavioral item (overly dependent). IC-TBI stress level only increased in one behavioral item (impulsivity). By injury severity, (i) mild (14.7%) showed no significant differences in emotional and behavioral either total-HIBS (ii) moderate (28.7%) showed significant emotional differences in two items (frequent complaining and mood change) and (iii) severe (56.6%) showed significant differences in emotional (often disputes topics) and behavioral (impulsivity) items. CONCLUSIONS: Our results suggest specific items in which informal caregivers could be supported considering cause or severity of the recipients of care.

The impact of COVID-19 on home, social, and productivity integration of people with chronic traumatic brain injury or stroke living in the community.

ABSTRACT: Compare community integration of people with stroke or traumatic brain injury (TBI) living in the community before and during the coronavirus severe acute respiratory syndrome coronavirus 2 disease (COVID-19) when stratifying by injury: participants with stroke (G1) and with TBI (G2); by functional independence in activities of daily living: independent (G3) and dependent (G4); by age: participants younger than 54 (G5) and older than 54 (G6); and by gender: female (G7) and male (G8) participants.Prospective observational cohort studyIn-person follow-up visits (before COVID-19 outbreak) to a rehabilitation hospital in Spain and on-line during COVID-19.Community dwelling adults (≥18 years) with chronic stroke or TBI.Community integration questionnaire (CIQ) the total-CIQ as well as the subscale domains (ie, home-CIQ, social-CIQ, productivity CIQ) were compared before and during COVID-19 using the Wilcoxon ranked test or paired t test when appropriate reporting Cohen effect sizes (d). The functional independence measure was used to assess functional independence in activities of daily living.Two hundred four participants, 51.4% with stroke and 48.6% with TBI assessed on-line between June 2020 and April 2021 were compared to their own in-person assessments performed before COVID-19.When analyzing total-CIQ, G1 (d = -0.231), G2 (d = -0.240), G3 (d = -0.285), G5 (d = -0.276), G6 (d = -0.199), G7 (d = -0.245), and G8 (d = -0.210) significantly decreased their scores during COVID-19, meanwhile G4 was the only group with no significant differences before and during COVID-19.In productivity-CIQ, G1 (d = -0.197), G4 (d = -0.215), G6 (d = -0.300), and G8 (d = -0.210) significantly increased their scores, meanwhile no significant differences were observed in G2, G3, G5, and G7.In social-CIQ, all groups significantly decreased their scores: G1 (d = -0.348), G2 (d = -0.372), G3 (d = -0.437), G4 (d = -0.253), G5 (d = -0.394), G6 (d = -0.319), G7 (d = -0.355), and G8 (d = -0.365).In home-CIQ only G6 (d = -0.229) significantly decreased, no significant differences were observed in any of the other groups.The largest effect sizes were observed in total-CIQ for G3, in productivity-CIQ for G6, in social-CIQ for G3 and in home-CIQ for G6 (medium effect sizes).Stratifying participants by injury, functionality, age or gender allowed identifying specific CIQ subtotals where remote support may be provided addressing them.

Long-term trajectories of motor functional independence after ischemic stroke in young adults: Identification and characterization using inpatient baseline assessments.

BACKGROUND: Stroke is a major worldwide cause of serious long-term disability. Most previous studies addressing functional independence included only inpatients with limited follow-up. OBJECTIVE: To identify novel classes of patients having similar temporal patterns in motor functional independence and relate them to baseline clinical features. METHODS: Retrospective observational cohort study, data were obtained for n = 428 adult patients with ischemic stroke admitted to rehabilitation (March 2005-March 2020), including baseline clinical features and follow-ups of motor Functional Independence Measure (mFIM) categorized as poor, fair or good. Growth mixture models (GMMs) were fitted to identify classes of patients with similar mFIM trajectories. RESULTS: GMM identified three classes of trajectories (1,664 mFIM assessments):C1 (11.2 %), 97.9% having poor admission mFIM, at 4.93 years 61.1% still poor, with the largest percentage of hypertension, neglect, dysphagia, diabetes and dyslipidemia of all three classes.C2 (23.1%), 99% had poor admission mFIM, 25% poor discharge mFIM, the largest percentage of aphasia and greatest mFIM gain, at 4.93 years only 6.2% still poor.C3 (65.7%) the youngest, lowest NIHSS, 37.7% poor admission mFIM, 73% good discharge mFIM, only 4.6% poor discharge mFIM, 90% good at 4.93 years. CONCLUSIONS: GMM identified novel motor functional classes characterized by baseline features.

To explain or not to explain?-Artificial intelligence explainability in clinical decision support systems.

Explainability for artificial intelligence (AI) in medicine is a hotly debated topic. Our paper presents a review of the key arguments in favor and against explainability for AI-powered Clinical Decision Support System (CDSS) applied to a concrete use case, namely an AI-powered CDSS currently used in the emergency call setting to identify patients with life-threatening cardiac arrest. More specifically, we performed a normative analysis using socio-technical scenarios to provide a nuanced account of the role of explainability for CDSSs for the concrete use case, allowing for abstractions to a more general level. Our analysis focused on three layers: technical considerations, human factors, and the designated system role in decision-making. Our findings suggest that whether explainability can provide added value to CDSS depends on several key questions: technical feasibility, the level of validation in case of explainable algorithms, the characteristics of the context in which the system is implemented, the designated role in the decision-making process, and the key user group(s). Thus, each CDSS will require an individualized assessment of explainability needs and we provide an example of how such an assessment could look like in practice.

Similar admission NIHSS may represent larger tissue-at-risk in patients with right-sided versus left-sided large vessel occlusion.

BACKGROUND: We investigated the effects of the side of large vessel occlusion (LVO) on post-thrombectomy infarct volume and clinical outcome with regard to admission National Institutes of Health Stroke Scale (NIHSS) score. METHODS: We retrospectively identified patients with anterior LVO who received endovascular thrombectomy and follow-up MRI. Applying voxel-wise general linear models and multivariate analysis, we assessed the effects of occlusion side, admission NIHSS, and post-thrombectomy reperfusion (modified Thrombolysis in Cerebral Infarction, mTICI) on final infarct distribution and volume as well as discharge modified Rankin Scale (mRS) score. RESULTS: We included 469 patients, 254 with left-sided and 215 with right-sided LVO. Admission NIHSS was higher in those with left-sided LVO (median (IQR) 16 (10-22)) than in those with right-sided LVO (14 (8-16), p>0.001). In voxel-wise analysis, worse post-thrombectomy reperfusion, lower admission NIHSS score, and poor discharge outcome were associated with right-hemispheric infarct lesions. In multivariate analysis, right-sided LVO was an independent predictor of larger final infarct volume (p=0.003). There was a significant three-way interaction between admission stroke severity (based on NIHSS), LVO side, and mTICI with regard to final infarct volume (p=0.041). Specifically, in patients with moderate stroke (NIHSS 6-15), incomplete reperfusion (mTICI 0-2b) was associated with larger final infarct volume (p<0.001) and worse discharge outcome (p=0.02) in right-sided compared with left-sided LVO. CONCLUSIONS: When adjusted for admission NIHSS, worse post-thrombectomy reperfusion is associated with larger infarct volume and worse discharge outcome in right-sided versus left-sided LVO. This may represent larger tissue-at-risk in patients with right-sided LVO when applying admission NIHSS as a clinical biomarker for penumbra.

Predictive models for independence after stroke rehabilitation: Maugeri external validation and development of a new model.

BACKGROUND: Many efforts have been devoted to identify predictors of functional outcomes after stroke rehabilitation. Though extensively recommended, there are very few external validation studies. OBJECTIVE: To externally validate two predictive models (Maugeri model 1 and model 2) and to develop a new model (model 3) that estimate the probability of achieving improvement in physical functioning (primary outcome) and a level of independence requiring no more than supervision (secondary outcome) after stroke rehabilitation. METHODS: We used multivariable logistic regression analysis for validation and development. Main outcome measures were: Functional Independence Measure (FIM) (primary outcome), Functional Independence Staging (FIS) (secondary outcome) and Minimal Clinically Important Difference (MCID). RESULTS: Patients with stroke admitted to a rehabilitation center from 2006 to 2019 were retrospectively studied (N = 710). Validation of Maugeri models confirmed very good discrimination: for model 1 AUC = 0.873 (0.833-0.915) and model 2 AUC = 0.803 (0.749-0.857). The Hosmer-Lemeshow χ2 was 6.07(p = 0.63) and 8.91(p = 0.34) respectively. Model 3 yielded an AUC = 0.894 (0.857-0.929) (primary outcome) and an AUC = 0.769 (0.714-0.825) (MCID). CONCLUSIONS: Discriminative power of both Maugeri models was externally confirmed (in a 20 years younger population) and a new model (incorporating aphasia) was developed outperforming Maugeri models in primary outcome and MCID.

On the usage of average Hausdorff distance for segmentation performance assessment: hidden error when used for ranking.

Average Hausdorff distance is a widely used performance measure to calculate the distance between two point sets. In medical image segmentation, it is used to compare ground truth images with segmentations allowing their ranking. We identified, however, ranking errors of average Hausdorff distance making it less suitable for applications in segmentation performance assessment. To mitigate this error, we present a modified calculation of this performance measure that we have coined "balanced average Hausdorff distance". To simulate segmentations for ranking, we manually created non-overlapping segmentation errors common in magnetic resonance angiography cerebral vessel segmentation as our use-case. Adding the created errors consecutively and randomly to the ground truth, we created sets of simulated segmentations with increasing number of errors. Each set of simulated segmentations was ranked using both performance measures. We calculated the Kendall rank correlation coefficient between the segmentation ranking and the number of errors in each simulated segmentation. The rankings produced by balanced average Hausdorff distance had a significantly higher median correlation (1.00) than those by average Hausdorff distance (0.89). In 200 total rankings, the former misranked 52 whilst the latter misranked 179 segmentations. Balanced average Hausdorff distance is more suitable for rankings and quality assessment of segmentations than average Hausdorff distance.

Toward Personalized Web-Based Cognitive Rehabilitation for Patients With Ischemic Stroke: Elo Rating Approach.

BACKGROUND: Stroke is a worldwide cause of disability; 40% of stroke survivors sustain cognitive impairments, most of them following inpatient rehabilitation at specialized clinical centers. Web-based cognitive rehabilitation tasks are extensively used in clinical settings. The impact of task execution depends on the ratio between the skills of the treated patient and the challenges imposed by the task itself. Thus, treatment personalization requires a trade-off between patients' skills and task difficulties, which is still an open issue. In this study, we propose Elo ratings to support clinicians in tasks assignations and representing patients' skills to optimize rehabilitation outcomes. OBJECTIVE: This study aims to stratify patients with ischemic stroke at an early stage of rehabilitation into three levels according to their Elo rating; to show the relationships between the Elo rating levels, task difficulty levels, and rehabilitation outcomes; and to determine if the Elo rating obtained at early stages of rehabilitation is a significant predictor of rehabilitation outcomes. METHODS: The PlayerRatings R library was used to obtain the Elo rating for each patient. Working memory was assessed using the DIGITS subtest of the Barcelona test, and the Rey Auditory Verbal Memory Test (RAVLT) was used to assess verbal memory. Three subtests of RAVLT were used: RAVLT learning (RAVLT075), free-recall memory (RAVLT015), and recognition (RAVLT015R). Memory predictors were identified using forward stepwise selection to add covariates to the models, which were evaluated by assessing discrimination using the area under the receiver operating characteristic curve (AUC) for logistic regressions and adjusted R2 for linear regressions. RESULTS: Three Elo levels (low, middle, and high) with the same number of patients (n=96) in each Elo group were obtained using the 50 initial task executions (from a total of 38,177) for N=288 adult patients consecutively admitted for inpatient rehabilitation in a clinical setting. The mid-Elo level showed the highest proportions of patients that improved in all four memory items: 56% (54/96) of them improved in DIGITS, 67% (64/96) in RAVLT075, 58% (56/96) in RAVLT015, and 53% (51/96) in RAVLT015R (P<.001). The proportions of patients from the mid-Elo level that performed tasks at difficulty levels 1, 2, and 3 were 32.1% (3997/12,449), 31.% (3997/12,449), and 36.9% (4595/12,449), respectively (P<.001), showing the highest match between skills (represented by Elo level) and task difficulties, considering the set of 38,177 task executions. Elo ratings were significant predictors in three of the four models and quasi-significant in the fourth. When predicting RAVLT075 and DIGITS at discharge, we obtained R2=0.54 and 0.43, respectively; meanwhile, we obtained AUC=0.73 (95% CI 0.64-0.82) and AUC=0.81 (95% CI 0.72-0.89) in RAVLT075 and DIGITS improvement predictions, respectively. CONCLUSIONS: Elo ratings can support clinicians in early rehabilitation stages in identifying cognitive profiles to be used for assigning task difficulty levels.

Opening the black box of artificial intelligence for clinical decision support: A study predicting stroke outcome.

State-of-the-art machine learning (ML) artificial intelligence methods are increasingly leveraged in clinical predictive modeling to provide clinical decision support systems to physicians. Modern ML approaches such as artificial neural networks (ANNs) and tree boosting often perform better than more traditional methods like logistic regression. On the other hand, these modern methods yield a limited understanding of the resulting predictions. However, in the medical domain, understanding of applied models is essential, in particular, when informing clinical decision support. Thus, in recent years, interpretability methods for modern ML methods have emerged to potentially allow explainable predictions paired with high performance. To our knowledge, we present in this work the first explainability comparison of two modern ML methods, tree boosting and multilayer perceptrons (MLPs), to traditional logistic regression methods using a stroke outcome prediction paradigm. Here, we used clinical features to predict a dichotomized 90 days post-stroke modified Rankin Scale (mRS) score. For interpretability, we evaluated clinical features' importance with regard to predictions using deep Taylor decomposition for MLP, Shapley values for tree boosting and model coefficients for logistic regression. With regard to performance as measured by Area under the Curve (AUC) values on the test dataset, all models performed comparably: Logistic regression AUCs were 0.83, 0.83, 0.81 for three different regularization schemes; tree boosting AUC was 0.81; MLP AUC was 0.83. Importantly, the interpretability analysis demonstrated consistent results across models by rating age and stroke severity consecutively amongst the most important predictive features. For less important features, some differences were observed between the methods. Our analysis suggests that modern machine learning methods can provide explainability which is compatible with domain knowledge interpretation and traditional method rankings. Future work should focus on replication of these findings in other datasets and further testing of different explainability methods.