Data-Driven Prognostication in Distal Medium Vessel Occlusions Using Explainable Machine Learning
=================================================================================================

* Mert Karabacak
* Burak Berksu Ozkara
* Tobias D. Faizy
* Trevor Hardigan
* Jeremy J. Heit
* Dhairya A. Lakhani
* Konstantinos Margetis
* J. Mocco
* Kambiz Nael
* Max Wintermark
* Vivek S. Yedavalli

## Abstract

**BACKGROUND AND PURPOSE:** Distal medium vessel occlusions (DMVOs) are estimated to cause acute ischemic stroke in 25%–40% of cases. Prognostic models can inform patient counseling and research by enabling outcome predictions. However, models designed specifically for DMVOs are lacking.

**MATERIALS AND METHODS:** This retrospective study developed a machine learning model to predict 90-day unfavorable outcome (defined as an mRS score of 3–6) in 164 patients with primary DMVO. A model developed with the TabPFN algorithm used selected clinical, laboratory, imaging, and treatment data with the least absolute shrinkage and selection operator feature selection. Performance was evaluated via 5-repeat 5-fold cross-validation. Model discrimination and calibration were evaluated. SHapley Additive Explanations (SHAP) identified influential features. A Web application deployed the model for individualized predictions.

**RESULTS:** The model achieved an area under the receiver operating characteristic curve of 0.815 (95% CI, 0.79–0.841) for predicting unfavorable outcome, demonstrating good discrimination, and a Brier score of 0.19 (95% CI, 0.177–0.202), demonstrating good calibration. SHAP analysis ranked admission NIHSS score, premorbid mRS, type of thrombectomy, modified TICI score, and history of malignancy as top predictors. The Web application enables individualized prognostication.

**CONCLUSIONS:** Our machine learning model demonstrated good discrimination and calibration for predicting 90-day unfavorable outcomes in primary DMVO strokes. This study demonstrates the potential for personalized prognostic counseling and research to support precision medicine in stroke care and recovery.

## ABBREVIATIONS:

AIS
:   acute ischemic stroke
AUPRC
:   area under the PRC
AUROC
:   area under the ROC curve
DMVO
:   distal medium vessel occlusion
ER
:   emergency department
IVT
:   IV thrombolysis
kNN
:   k-nearest neighbor
LASSO
:   least absolute shrinkage and selection operator
mTICI
:   modified TICI
PDP
:   partial dependence plot
PRC
:   precision-recall curve
ROC
:   receiver operating characteristic
SHAP
:   SHapley Additive ExPlanations
ST
:   stroke thrombectomy

SUMMARY

#### PREVIOUS LITERATURE:

DMVOs account for a considerable portion of acute ischemic strokes. While prognostic models for large vessel occlusion stroke outcomes exist, those specifically designed for DMVOs are lacking. Previous studies have explored machine learning approaches for predicting outcomes in acute ischemic stroke, but these models are not tailored to the unique characteristics of DMVOs. One prior study developed machine learning models to predict an NIHSS shift in patients with DMVO but did not address longer-term functional outcomes measured by the mRS.

#### KEY FINDINGS:

Our machine learning model, using the TabPFN algorithm, achieved good discrimination with an area under the receiver operating characteristic curve of 0.815 (95% CI, 0.79–0.841) and calibration with a Brier score of 0.19 (95% CI, 0.177–0.202) in predicting 90-day unfavorable outcomes in patients with DMVO. The model identified admission NIHSS score, premorbid mRS, thrombectomy type, modified TICI score, and malignancy history as top predictors.

#### KNOWLEDGE ADVANCEMENT:

This study presents the first prognostic machine learning model specifically designed for predicting mRS outcomes in patients with DMVO. By integrating clinical, laboratory, imaging, and treatment data, our model provides a tool for personalized prognostication in this important stroke subtype, potentially informing clinical decision-making and research strategies.

Distal medium vessel occlusions (DMVOs), also known as medium vessel occlusions, are estimated to cause acute ischemic stroke (AIS) in 25%–40% of cases.1 They are most commonly defined as occlusions in non-co-dominant M2, M3, or M4 segments of the MCA, anterior cerebral artery, posterior cerebral artery, posterior inferior cerebellar artery, superior cerebellar artery, or anterior inferior cerebellar artery.2 The clinical manifestations of DMVOs are heterogeneous, and the optimal imaging technique for diagnosis is yet to be defined.3,4 DMVOs represent an emerging application for stroke thrombectomy (ST); thus, it is an active area of research.1,5 While ST has begun to be used for DMVOs in clinical practice, a better understanding of the efficacy and safety of DMVOs in ST is necessary, because findings from retrospective studies and meta-analyses have been conflicting.6

These challenges underscore the need for accurate prognostic models to anticipate disease progression and outcomes in patients with DMVOs, regardless of whether they receive ST, medical therapy, or both. While prognostic scales, nomograms, and machine learning approaches have been extensively studied for predicting outcomes in AIS, models tailored specifically for DMVOs are lacking in the literature.7⇓-9 Machine learning models to predict the NIHSS shift of patients with DMVOs were developed in a previously published study.10 In the current study, we aimed to address the need for accurate individual-level predictions of longer-term functional outcomes in patients with primary DMVO. Therefore, we developed a novel model based on a modified Prior-Data Fitted Network architecture that leverages clinical, laboratory, imaging, and treatment variables to predict unfavorable outcome, defined as an mRS score of 3–6, a standard measure in clinical trials.11

## MATERIALS AND METHODS

This was a retrospective cohort study using machine learning to predict unfavorable functional outcome, defined as an mRS score of 3–6 at 90 days, in patients with primary DMVO. Patients were dichotomized into 2 outcome groups [favorable (mRS 0–2) versus unfavorable (mRS 3–6)], and a binary classifier was developed to predict the outcome of interest.11 The study adhered to the Transparent Reporting of Multivariable Prediction Models for Individual Prognosis or Diagnosis–Artificial Intelligence (TRIPOD+AI).12 The data-processing pipeline is depicted in Fig 1 as a summary of our methodology.

![FIG 1.](http://www.ajnr.org/https://ajnr-sso.highwirestaging.com/content/ajnr/46/4/725/F1.medium.gif)

[FIG 1.](http://www.ajnr.org/content/46/4/725/F1)

FIG 1. 
Data-processing pipeline. MCC indicates Matthews Correlation Coefficient; min, minimum; max, maximum.

### Ethics Approval

This study was performed in accordance with the Helsinki Declaration (as revised in 2013). The Johns Hopkins Hospital institutional review board approved the study. The requirement for individual informed consent was waived due to the retrospective nature of the study.

### Study Population

We used data from 2 comprehensive stroke centers, Johns Hopkins Hospital and Johns Hopkins Bayview Medical Center. Consecutive patients admitted between January 1, 2017, and October 16, 2022, were screened for eligibility. A DMVO was defined as an arterial occlusion involving the anterior cerebral artery, MCA M2–M4, posterior cerebral artery, posterior inferior cerebellar artery, anterior inferior cerebellar artery, or superior cerebellar artery.2 The diagnosis of AIS was made on the basis of clinical examination and was confirmed with CT or MR imaging of the brain. Patients were included if they met the following criteria: 1) admission within 24 hours of symptom onset; 2) 18 years of age or older; and 3) confirmed primary diagnosis of DMVO using CTA or CTP. Patients were excluded if outcome data were incomplete or if the DMVO was secondary to an iatrogenic embolus from the endovascular treatment of a different occlusion.

### Demographic and Clinical Data

Demographic and clinical data were retrospectively extracted from the electronic medical records. We collected the following variables: age, sex, race, smoking status, comorbidities (diabetes, dyslipidemia, hypertension, heart disease, atrial fibrillation, chronic kidney disease, sleep apnea, peripheral vascular disease), prior deep vein thrombosis or pulmonary emboli, prior stroke or TIA, history of malignancy, antiplatelet use, body mass index, admission vitals (systolic blood pressure, diastolic blood pressure, heart rate, respiratory rate, and oxygen saturation), admission NIHSS score, premorbid (prestroke) mRS score, stroke etiology, and 90-day mRS score.

### Laboratory Data

Peripheral venous blood samples were collected from all patients on arrival at the emergency department (ER) per institutional standard stroke protocol. Samples were processed and analyzed uniformly using the same methods at the clinical laboratories of the 2 hospitals. We retrospectively retrieved the following admission laboratory parameters: electrolytes (sodium, potassium, chloride, calcium, phosphorus, magnesium), carbon dioxide, glucose, blood urea nitrogen, creatinine, albumin, total protein, liver tests (total bilirubin, alkaline phosphatase, alanine transaminase, aspartate aminotransferase), complete blood count (hematocrit, hemoglobin, white blood cell count, platelet count), and coagulation studies (partial thromboplastin time, international normalized ratio).

### Imaging Data

Both centers used helical scanners from Somatom Flash and/or Drive (Siemens) to perform comprehensive baseline CT imaging. The imaging parameters applied in this study aligned with those detailed in an earlier published study.13

Noncontrast CTs, CTAs, and CTPs of all patients were evaluated by a board-certified neuroradiologist (V.S.Y., with 9 years of experience in neuroradiology), and imaging notes were used to collect imaging data. RapidAI (iSchemaView) was used to assist in the interpretation of CTP findings. This assessment was carried out in tandem with examining all available imaging and clinical data. The same neuroradiologist also confirmed and gathered the presence of any DMVO, the baseline NCCT-ASPECTS, the occluded vessel, the laterality of the occlusion, the MCA dot sign, and the occurrence of hemorrhagic transformation.

### Treatment Data

All thrombectomy procedures were performed by 1 of 4 credentialed interventional neuroradiologists or endovascular neurosurgeons. Device selection was at the discretion of the operator and limited to Food and Drug Administration–cleared options. We collected the following treatment variables: IV thrombolysis (IVT) administration, stroke thrombectomy performance, thrombectomy technique utilized, final reperfusion grade per modified TICI (mTICI) score as assessed by the interventionist, and the number of thrombectomy passes. Additionally, we recorded the following time intervals (measured in minutes): from groin puncture to first pass, groin puncture to recanalization, first pass to recanalization, last known well to ER arrival, symptom onset to ER arrival, ER to CT scan, last known well to CT, ER to groin puncture, ER to IVT bolus, and ER to final recanalization.

### Data Preprocessing and Feature Selection

To prevent exclusion bias, we used imputation methods for missing data. Of the continuous variables, 25 had at least 1 missing value. After excluding 1 variable missing in >25% of patients (symptom onset to arrival at the ER), the k-nearest neighbor (kNN) algorithm (k = 5) imputed missing values by leveraging data from the whole data set.14 The kNN approach fills in missing data using values from the 5 most similar cases. For categoric variables, 7 had missing values. No variable was excluded because no variable was missing in >25% of patients, and the missing values were imputed using the mode.

Feature selection was performed to determine the variables most relevant for predicting outcomes from the preprocessed data set. The least absolute shrinkage and selection operator (LASSO) regression algorithm (α = .01) was used for this purpose.15 LASSO performs both variable selection and regularization to improve prediction accuracy. It adds a penalty proportional to the absolute size of the model coefficients. The degree of penalty is controlled by a tuning parameter λ. As λ increases, more coefficients are shrunk toward zero, effectively removing less predictive features.

We implemented LASSO feature selection within each cross-validation fold for which the details are given below. First, the training data in each fold were minimum-maximum scaled to normalize features. A LASSO model was then fitted to the scaled training data in each fold to select impactful features with nonzero model coefficients. The selected features were recorded per fold. Finally, features chosen in >50% of folds were selected as the final input variables to reduce variability.

### Model Development and Evaluation

We used TabPFN, a modified Tabular Prior-Data Fitted Network architecture, for model development. TabPFN uses a meta-learning framework to enable adaptation to new, unseen data by learning from diverse data sets.16 Prior-data fitted networks like TabPFN are pretrained on synthetic data to approximate the Bayesian inference on real-world data.16 The pretraining enables TabPFN to capture complex patterns in tabular data and transition smoothly to new data sets.

Model performance was evaluated using a 5-repeat 5-fold stratified cross-validation framework. In each repeat, the data were split into 5 roughly equal folds with a different random split, balancing outcome class ratios (stratification) to guarantee class balance across folds. The use of 5 repeats in this framework serves to increase the robustness of our performance estimates, reduce the impact of any particular random data split, mitigate potential overfitting to a single partition, and provide a more comprehensive assessment of model stability. This approach allows more reliable estimation of confidence intervals for our performance metrics. Within each fold during every repeat, the initial training set (80% of data) was further segmented into a final training subset (70% of full data) and a validation subset (10% of full data). This step resulted in a final 70:10:20 ratio for training to validation to hold-out testing. The validation subsets allowed sigmoid calibration to align predicted risks with observed outcomes. Model discrimination, calibration, and accuracy were then evaluated on the held-out test folds.

The calibrated TabPFN model generated predictions and probability estimates on each test fold across the 5 cross-validation repeats. Overall performance was evaluated by aggregating results across all folds and repeats. Cross-validation enabled reliable assessment of generalizable predictive performance. To improve interpretability, SHapley Additive ExPlanations (SHAP) were used to determine relative feature importance. The SHAP plot displayed selected features hierarchically, with the most influential at the top. Additionally, partial dependence plots (PDPs) showed the isolated effect of individual features on predicted outcomes. PDPs illustrate the isolated effect of a single feature on the predicted output of the model, revealing the extent to which individual features influence the predictions. The model code is available in the study GitHub repository ([https://github.com/mertkarabacak/DMVO-mRS](https://github.com/mertkarabacak/DMVO-mRS)) for full transparency.

Model performance was evaluated graphically using a receiver operating characteristic (ROC) curve, which displays the ability of a binary classifier to discriminate between positive and negative classes; a precision-recall curve (PRC) illustrating the trade-off between precision and recall; a calibration plot for visually assessing the agreement between predicted probabilities and observed outcomes; and a confusion matrix that aggregates predictions across all folds and repeats to show the number of true-positives, true-negatives, false-positives, and false-negatives. Numerically, we computed precision, recall, the F1-score, the Matthews Correlation Coefficient, the area under the ROC curve (AUROC), the area under the PRC (AUPRC), and the Brier score. A 95% CI for each metric was calculated using a bootstrap approach with 1000 resampled data sets. This calculation involved sampling with replacement from the original data set to generate 1000 new samples. The confidence interval was determined by finding the 2.5th and 97.5th percentiles of the bootstrapped metric mean values.

### Web Application

We developed a Web application to enable health care professionals and researchers to generate individualized predictions using our model. The application was deployed via Hugging Face, a platform for sharing machine learning models. Our implementation code is publicly available on the same platform for full transparency. The functionality of the Web application is demonstrated in the Supplementary Video 1. It can be accessed at the following URL: [https://huggingface.co/spaces/MSHS-Neurosurgery-Research/DMVO-mRS](https://huggingface.co/spaces/MSHS-Neurosurgery-Research/DMVO-mRS).

## RESULTS

Initially, 212 patients who met all the inclusion criteria were identified. Forty-eight were excluded due to missing 90-day mRS data, leaving 164 patients for analysis. The group with a favorable outcome (90-day mRS 0–2) included 90 patients, while the unfavorable outcome group (90-day mRS 3–6) comprised 74 patients. The median age was 71 years. ST was performed in 43 (47.8%) and 41 (55.4%) patients in the favorable and unfavorable outcome groups, respectively. IVT was administered to 34 (37.8%) patients and 22 (29.7%) patients in favorable and unfavorable outcome groups, respectively. The Supplemental Data summarize key cohort characteristics, including demographics and selected variables. The Supplemental Data provide full details on all baseline clinical, laboratory, and imaging parameters.

On executing the LASSO regression algorithm, we determined the following features (*n* = 16) to be the most pertinent for predicting the outcome of interest and used these for model development from the initial feature set (*n* = 64): age, current or former smoker status, diabetes, hypertension, deep vein thrombosis or pulmonary emboli, history of malignancy, antiplatelet use, admission hemoglobin, admission body mass index, admission NIHSS, premorbid mRS, the MCA dot sign, occlusion laterality, IVT, type of thrombectomy, and mTICI.

The model achieved a strong predictive performance with a precision of 0.711 (95% CI, 0.634–0.765), recall of 0.628 (95% CI, 0.553–0.694), F1-score of 0.656 (95% CI, 0.585–0.708), accuracy of 0.724 (95% CI, 0.696–0.752), a Matthews Correlation Coefficient of 0.45 (95% CI, 0.39–0.503), and a Brier score of 0.19 (95% CI, 0.177–0.202). The AUROC was 0.815 (95% CI, 0.79–0.841), indicating good discrimination. The AUPRC of 0.808 (95% CI, 0.781–0.832) also demonstrated strong precision-recall performance (Table). The model ROC curve, PRC, calibration curve, and confusion matrix are displayed in Fig 2. Figure 3 shows the relative feature importance by SHAP values. The top 5 predictors were admission NIHSS score, premorbid mRS score, type of thrombectomy, mTICI score, and history of malignancy. To demonstrate the isolated impact of key variables, PDPs (Supplemental Data) showed the effects for the 9 most influential features.

![FIG 2.](http://www.ajnr.org/https://ajnr-sso.highwirestaging.com/content/ajnr/46/4/725/F2.medium.gif)

[FIG 2.](http://www.ajnr.org/content/46/4/725/F2)

FIG 2. 
ROC (*A*), PRC (*B*), calibration curve (*C*), and confusion matrix (*D*) of the model.

![FIG 3.](http://www.ajnr.org/https://ajnr-sso.highwirestaging.com/content/ajnr/46/4/725/F3.medium.gif)

[FIG 3.](http://www.ajnr.org/content/46/4/725/F3)

FIG 3. 
SHAP plot of the model sorting features by their relative importance.

|        Performance Metric        | Metric Value (95% CI) |
|:--------------------------------:| --------------------- |
|            Precision             | 0.711 (0.634–0.765)   |
|              Recall              | 0.628 (0.553–0.694)   |
|             F1 Score             | 0.656 (0.585–0.708)   |
|             Accuracy             | 0.724 (0.696–0.752)   |
| Matthews Correlation Coefficient | 0.45 (0.390–0.503)    |
|              AUROC               | 0.815 (0.790–0.841)   |
|              AUPRC               | 0.808 (0.781–0.832)   |

Model performance

## DISCUSSION

While results from ongoing trials will be critical for determining the efficacy of ST for DMVOs, ST is already being used clinically for many patients with DMVO.1 Reliable prognostication tools may be especially valuable, given the current lack of consensus around optimal DMVO treatment approaches. These tools facilitate several applications: enabling informed discussions around likely prognosis, serving as a quality check to prompt protocol reassessment when outcomes are worse than expected, and streamlining patient stratification for research and clinical trials.

This study demonstrates the potential of machine learning models to improve prognostic predictions for patients with DMVO by developing a practical tool to forecast unfavorable functional outcomes at 90 days. A unique aspect was integrating the model into a user-friendly Web application to provide clinicians with personalized prognostic assessments. Our model achieved an AUROC of 0.815 for predicting unfavorable outcome (mRS score of 3–6) regardless of treatment approach. Additionally, the model demonstrated good calibration, with a Brier score of 0.19 and a near-ideal calibration curve (Fig 2C). With these promising discrimination and calibration results, our study showed that machine learning could enable valuable individualized prognostication for DMVOs using key clinical and imaging parameters. Moreover, to our knowledge based on the literature review, this result represents the first prognostic model specifically designed for predicting mRS outcomes in a DMVO population.

Our methodology enables precise outcome predictions for patients with DMVO while also improving the interpretability of those forecasts. The SHAP feature importance plot (Fig 3) provides a global explanation of overall model behavior by revealing general patterns of how key variables relate to outcomes across the full data set. Additionally, the SHAP plots integrated into our Web application deliver local explanations that give granular insight into how unique predictions are impacted by certain variables in specific cases. This functionality allows a personalized understanding of the drivers of an individual prediction, which has not been readily accessible in most prior models. The local SHAP visualizations not only enhance interpretability but also bolster the credibility of our model when combined with clinical judgment. By enabling clinicians to review the variables influencing each prediction, SHAP plots facilitate expert evaluation of model behavior and output. This synergy between data-driven insight and human expertise can improve the acceptance of model-based predictions, underscoring the potential of the approach to meaningfully inform prognostication.

In our study, we used LASSO feature selection within each cross-validation fold, effectively reducing the number of covariates by keeping only those with high predictive power. This rigorous selection process resulted in fewer covariates, emphasizing model accuracy and stability over complexity. According to the global explanations provided by the SHAP analysis, the admission NIHSS score was the most important feature of our model. The NIHSS is a quantitative assessment of neurologic deficits associated with stroke that has been shown to be reliable within and between raters and to have significant predictive power for functional outcomes.17 Prior research indicates that patients with very high or very low NIHSS scores often have predictable trajectories, with low scores suggesting probable good recovery and high scores indicating likely poor outcomes. Patients at the extreme ends are also less likely to have a large observable treatment effect.18,19 Although some prognostic scales use the NIHSS score as a predictor, none have been designed specifically for DMVOs.7,8 Aligning with prior work, the admission NIHSS score was informative for predicting 90-day functional status in our DMVO cohort. Because DMVOs represent an AIS subtype, this logical association is consistent and emphasizes potential generalizability. Prestroke functional status, assessed by the premorbid mRS score, also strongly contributed to predictions. Previous research suggests that higher premorbid disability levels are associated with worse outcomes and mortality, even for patients receiving ST.20⇓⇓-23 Similarly, in our cohort, 6 of 7 patients with premorbid mRS scores of >3, indicating significant prior disability, had died by the 90-day follow-up, versus only 30 deaths among 156 patients with premorbid mRS scores of ≤3. Therefore, the baseline functionality level provided a useful predictive signal regarding the probability of favorable recovery in our study.

The type of ST technique used ranked as the third most important predictor in our model. A recent systematic review found heterogeneity among ongoing DMVO trials regarding the thrombectomy approach: Some studies mandated stent retrievers while others allowed operator discretion across techniques.5 These studies may provide insight into how and why the specific thrombectomy technique impacts outcomes and contributes significant predictive signal, as seen in the feature importance ranking of our model. The final reperfusion grade, as assessed by the mTICI score, also featured prominently as the fourth most influential factor. The mTICI scale represents an enhancement from the original TICI system designed specifically for cerebral circulation.24 It is the primary scale recommended for evaluating reperfusion therapy in patients undergoing ST because it is tailored for cerebral circulation, has high interrater reliability, and is a strong predictor of clinical outcomes.25 Regarding the prognostic impact of the mTICI score in patients with DMVOs treated with ST, some studies have shown its positive effect on favorable outcomes, stressing the importance of the mTICI score as a critical variable in our model.26,27 This clarification helps contextualize why the mTICI score would logically serve as an informative predictor. The history of malignancy and diabetes mellitus was another important feature revealed by SHAP analysis. It has been shown that systemic malignancy is linked to an increased risk of ischemic stroke, and patients with a history of cancer are more likely to have recurrent strokes and die from cardiovascular disease.28,29 Furthermore, diabetes has long been recognized as a risk factor for stroke mortality.30 These risk factors provide further context as to why a history of malignancy and diabetes would logically be an informative predictor.

Our study has several limitations that should be considered when interpreting the results. The retrospective design and modest cohort size (*n* = 164) inherently limit the generalizability of our findings. While we rigorously evaluated model performance using a 5-repeat, 5-fold stratified cross-validation methodology, this approach does not replace the value of external validation in entirely separate populations. The absence of an independent test set further limits our ability to fully assess the generalizability of the model. Future studies should aim for larger, multicenter, prospective cohorts with independent test sets for more robust validation. Data were drawn from 2 high-volume comprehensive stroke centers, which may not represent the full spectrum of clinical settings where DMVOs are treated. Outcomes may depend substantially on facility volume, operator experience, and other institutional factors. The evaluation of imaging data by a single neuroradiologist and the involvement of only 4 neurointerventionalists introduce potential bias, particularly given the importance of thrombectomy technique as a predictive factor.

We acknowledge limitations in our input variables. The NIHSS may not fully capture stroke symptoms, especially in posterior circulation strokes. The mTICI score, designed for anterior circulation large vessel occlusions, may not be optimal for all vessels involved in DMVOs. The mRS, our outcome measure, has known variability within each category. The use of kNN for imputing missing data has limitations in dealing with nonrandom missingness, potentially introducing bias. Additionally, a substantial number of patients were excluded due to missing 90-day mRS scores, which could affect the generalizability of the model. Last, while certain variables demonstrate prognostic relationships with outcomes, these associations should not be inferred as causal without further analysis. The model predicts outcomes based on patterns and correlations without determining causality. Model outputs should be viewed as prognostic forecasts rather than endorsements of causal mechanisms or treatment effectiveness.

Future studies should aim to address these limitations through larger, prospective, multicenter designs with more diverse patient populations and clinical settings. Incorporating more granular details, including but not limited to discharge disposition, discharge mRS, use of other functional neurorehabilitation scales, and readmissions within 30 days, may provide even more predictive models in the future. No clinical decisions or interventions should be actively guided by this model without further validation, because its current function is exclusively prognostic rather than prescriptive.

## CONCLUSIONS

Identifying outcome risks and engaging in shared decision-making is crucial for patient care in DMVOs. Our machine learning approach, using the TabPFN algorithm, achieved good discrimination and calibration in predicting 90-day functional outcomes in patients with DMVO. By translating predictive modeling into accessible and interpretable risk estimates, our methodology exemplifies how precision medicine tools can empower granular prognosis for stroke variants like DMVOs. External validation with larger multicenter data sets is needed to confirm generalizability before integrating such tools into routine practice.

## Footnotes

*   M. Karabacak and B.B. Ozkara contributed equally to this work.

*   [Disclosure forms](https://www.ajnr.org/sites/default/files/additional-assets/Disclosures/April%202025/0783.pdf) provided by the authors are available with the full text and PDF of this article at [www.ajnr.org](http://www.ajnr.org).

## References

1.  1.Kobeissi H, Bilgin C, Ghozy S, et al. A review of acute ischemic stroke caused by distal, medium vessel occlusions. Interv Neuroradiol 2023 Aug 29 [Epub ahead of print] doi:10.1177/15910199231197616 pmid:37644821
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1177/15910199231197616&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=37644821&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

2.  2.Saver JL, Chapot R, Agid R, et al; Distal Thrombectomy Summit Group. Thrombectomy for distal, medium vessel occlusions: a consensus statement on present knowledge and promising directions. Stroke 2020;51:2872–84 doi:10.1161/STROKEAHA.120.028956 pmid:32757757
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1161/STROKEAHA.120.028956&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=32757757&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

3.  3.Sousa JA, Sondermann A, Bernardo-Castro S, et al. CTA and CTP for detecting distal medium vessel occlusions: a systematic review and meta-analysis. AJNR Am J Neuroradiol 2023;45:51–56 doi:10.3174/ajnr.A8080 pmid:38164544
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.3174/ajnr.A8080&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=38164544&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

4.  4.Sousa JA, Sondermann A, Bernardo-Castro S, et al. Diagnostic accuracy of CT angiography and CT perfusion imaging for detecting distal medium vessel occlusions: protocol for a systematic review and meta-analysis. PLoS One 2023;18:e0284116 doi:10.1371/journal.pone.0284116 pmid:37036841
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1371/journal.pone.0284116&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=37036841&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

5.  5.Bilgin C, Bolsegui ML, Ghozy S, et al. Common design and data elements reported in active mechanical thrombectomy trials focusing on distal medium vessel occlusions and minor strokes: a systematic review. J Neurointerv Surg 2024 Jan 11 [Epub ahead of print] doi:10.1136/jnis-2023-021073 pmid:38212110
    
    [Abstract/FREE Full Text](http://www.ajnr.org/lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6MTE6Im5ldXJpbnRzdXJnIjtzOjU6InJlc2lkIjtzOjE4OiJqbmlzLTIwMjMtMDIxMDczdjEiO3M6NDoiYXRvbSI7czoxOToiL2FqbnIvNDYvNC83MjUuYXRvbSI7fXM6ODoiZnJhZ21lbnQiO3M6MDoiIjt9) 

6.  6.Nedelcu S, Gulati A, Henninger N. Endovascular therapy versus medical management for mild strokes due to medium and distal vessel occlusions. Interv Neuroradiol 2023 Dec 17 [Epub ahead of print] doi:10.1177/15910199231216510 pmid:38105534
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1177/15910199231216510&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=38105534&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

7.  7.Matsumoto K, Nohara Y, Soejima H, et al. Stroke prognostic scores and data-driven prediction of clinical outcomes after acute ischemic stroke. Stroke 2020;51:1477–83 doi:10.1161/STROKEAHA.119.027300 pmid:32208843
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1161/STROKEAHA.119.027300&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=32208843&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

8.  8.Drozdowska BA, Singh S, Quinn TJ. Thinking about the future: a review of prognostic scales used in acute stroke. Front Neurol 2019;10:274 doi:10.3389/fneur.2019.00274 pmid:30949127
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.3389/fneur.2019.00274&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=30949127&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

9.  9.Mainali S, Darsie ME, Smetana KS. Machine learning in action: stroke diagnosis and outcome prediction. Front Neurol 2021;12:734345 doi:10.3389/fneur.2021.734345 pmid:34938254
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.3389/fneur.2021.734345&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=34938254&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

10. 10.Ozkara BB, Karabacak M, Kotha A, et al. Development of machine learning models for predicting outcome in patients with distal medium vessel occlusions: a retrospective study. Quant Imaging Med Surg 2023;13:5815–30 doi:10.21037/qims-23-154 pmid:37711830
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.21037/qims-23-154&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=37711830&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

11. 11.Weisscher N, Vermeulen M, Roos YB, et al. What should be defined as good outcome in stroke trials; a modified Rankin score of 0-1 or 0-2? J Neurol 2008;255:867–74 doi:10.1007/s00415-008-0796-8 pmid:18338195
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1007/s00415-008-0796-8&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=18338195&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 
    
    [Web of Science](http://www.ajnr.org/lookup/external-ref?access_num=000257629900014&link_type=ISI) 

12. 12.Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 2024;385:e078378 doi:10.1136/bmj-2023-078378 pmid:38626948
    
    [FREE Full Text](http://www.ajnr.org/lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiRlVMTCI7czoxMToiam91cm5hbENvZGUiO3M6MzoiYm1qIjtzOjU6InJlc2lkIjtzOjIwOiIzODUvYXByMTZfMTEvZTA3ODM3OCI7czo0OiJhdG9tIjtzOjE5OiIvYWpuci80Ni80LzcyNS5hdG9tIjt9czo4OiJmcmFnbWVudCI7czowOiIiO30=) 

13. 13.Ozkara BB, Karabacak M, Hamam O, et al. Prediction of functional outcome in stroke patients with proximal middle cerebral artery occlusions using machine learning models. J Clin Med 2023;12:839 doi:10.3390/jcm12030839 pmid:36769491
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.3390/jcm12030839&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=36769491&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

14. 14.Beretta L, Santaniello A. Nearest neighbor imputation algorithms: a critical evaluation. BMC Med Inform Decis Mak 2016;16(Suppl 3):74 doi:10.1186/s12911-016-0318-z pmid:27454392
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1186/s12911-016-0318-z&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=27454392&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

15. 15.Tibshirani R. Regression shrinkage and selection via the Lasso. Journal of the Royal Statistical Society Series B (Methodological) 1996;58:267–88 doi:10.1111/j.2517-6161.1996.tb02080.x
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1111/J.2517-6161.1996.TB02080.X&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=WOS&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 
    
    [Web of Science](http://www.ajnr.org/lookup/external-ref?access_num=A1996TU31400017&link_type=ISI) 

16. 16.Hollmann N, Müller S, Eggensperger K, et al. TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second. arXiv 2022 [https://doi.org/10.48550/ARXIV.2207.01848](https://doi.org/10.48550/ARXIV.2207.01848). Accessed March 10, 2024
    
    

17. 17.Saber H, Saver JL. Distributional validity and prognostic power of the National Institutes of Health Stroke Scale in US administrative claims data. JAMA Neurol 2020;77:606–12 doi:10.1001/jamaneurol.2019.5061 pmid:32065612
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1001/jamaneurol.2019.5061&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=32065612&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

18. 18.Saver JL. Optimal endpoints for acute stroke therapy trials: best ways to measure treatment effects of drugs and devices. Stroke 2011;42:2356–62 doi:10.1161/STROKEAHA.111.619122 pmid:21719772
    
    [Abstract/FREE Full Text](http://www.ajnr.org/lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6OToic3Ryb2tlYWhhIjtzOjU6InJlc2lkIjtzOjk6IjQyLzgvMjM1NiI7czo0OiJhdG9tIjtzOjE5OiIvYWpuci80Ni80LzcyNS5hdG9tIjt9czo4OiJmcmFnbWVudCI7czowOiIiO30=) 

19. 19.Ntaios G, Faouzi M, Ferrari J, et al. An integer-based score to predict functional outcome in acute ischemic stroke: the ASTRAL score. Neurology 2012;78:1916–22 doi:10.1212/WNL.0b013e318259e221 pmid:22649218
    
    [Abstract/FREE Full Text](http://www.ajnr.org/lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6OToibmV1cm9sb2d5IjtzOjU6InJlc2lkIjtzOjEwOiI3OC8yNC8xOTE2IjtzOjQ6ImF0b20iO3M6MTk6Ii9ham5yLzQ2LzQvNzI1LmF0b20iO31zOjg6ImZyYWdtZW50IjtzOjA6IiI7fQ==) 

20. 20.de Havenon A, Castonguay A, Nogueira R, et al. Prestroke disability and outcome after thrombectomy for emergent anterior circulation large vessel occlusion stroke. Neurology 2021;97:e1914–19 doi:10.1212/WNL.0000000000012827 pmid:34544817
    
    [Abstract/FREE Full Text](http://www.ajnr.org/lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6OToibmV1cm9sb2d5IjtzOjU6InJlc2lkIjtzOjExOiI5Ny8xOS9lMTkxNCI7czo0OiJhdG9tIjtzOjE5OiIvYWpuci80Ni80LzcyNS5hdG9tIjt9czo4OiJmcmFnbWVudCI7czowOiIiO30=) 

21. 21.Kwok CS, Clark A, Ford GA, et al. Association between prestroke disability and inpatient mortality and length of acute hospital stay after acute stroke. J Am Geriatr Soc 2012;60:726–32 doi:10.1111/j.1532-5415.2011.03889.x pmid:22316344
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1111/j.1532-5415.2011.03889.x&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=22316344&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

22. 22.Leker RR, Gavriliuc P, Yaghmour NE, et al. Increased risk for unfavorable outcome in patients with pre-existing disability undergoing endovascular therapy. J Stroke Cerebrovasc Dis 2018;27:92–96 doi:10.1016/j.jstrokecerebrovasdis.2017.08.007 pmid:28882658
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1016/j.jstrokecerebrovasdis.2017.08.007&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=28882658&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

23. 23.Salwi S, Cutting S, Salgado AD, et al. Mechanical thrombectomy in patients with ischemic stroke with prestroke disability. Stroke 2020;51:1539–45 doi:10.1161/STROKEAHA.119.028246 pmid:32268851
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1161/strokeaha.119.028246&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=32268851&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

24. 24.Zaidat OO, Yoo AJ, Khatri P, et al; STIR Thrombolysis in Cerebral Infarction (TICI) Task Force. Recommendations on angiographic revascularization grading standards for acute ischemic. Stroke 2013;44:2650–63 doi:10.1161/STROKEAHA.113.001972 pmid:23920012
    
    [FREE Full Text](http://www.ajnr.org/lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiRlVMTCI7czoxMToiam91cm5hbENvZGUiO3M6OToic3Ryb2tlYWhhIjtzOjU6InJlc2lkIjtzOjk6IjQ0LzkvMjY1MCI7czo0OiJhdG9tIjtzOjE5OiIvYWpuci80Ni80LzcyNS5hdG9tIjt9czo4OiJmcmFnbWVudCI7czowOiIiO30=) 

25. 25.Chamorro Á, Blasco J, López A, et al. Complete reperfusion is required for maximal benefits of mechanical thrombectomy in stroke patients. Sci Rep 2017;7:11636 doi:10.1038/s41598-017-11946-y pmid:28912596
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1038/s41598-017-11946-y&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=28912596&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

26. 26.Hulscher F, Farouki Y, Mine B, et al. Predictors of good clinical outcome after thrombectomy for distal medium vessel occlusions. World Neurosurg 2022;160:e566–72 doi:10.1016/j.wneu.2022.01.067 pmid:35077884
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1016/j.wneu.2022.01.067&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=35077884&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

27. 27.Abdelrady M, Derraz I, Dargazanli C, et al. Complete recanalization predicts favorable outcome in patients with distal M2-M3 middle cerebral artery occlusions following endovascular thrombectomy. J Neuroradiol 2023;50:230–36 doi:10.1016/j.neurad.2022.11.007 pmid:36436611
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1016/j.neurad.2022.11.007&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=36436611&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

28. 28.Lau KK, Wong YK, Teo KC, et al. Stroke patients with a past history of cancer are at increased risk of recurrent stroke and cardiovascular mortality. PLoS One 2014;9:e88283 doi:10.1371/journal.pone.0088283 pmid:24523883
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1371/journal.pone.0088283&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=24523883&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

29. 29.Yoo J, Nam HS, Kim YD, et al. Short-term outcome of ischemic stroke patients with systemic malignancy. Stroke 2019;50:507–11 doi:10.1161/STROKEAHA.118.023044 pmid:30626288
    
    [CrossRef](http://www.ajnr.org/lookup/external-ref?access_num=10.1161/STROKEAHA.118.023044&link_type=DOI) 
    
    [PubMed](http://www.ajnr.org/lookup/external-ref?access_num=30626288&link_type=MED&atom=%2Fajnr%2F46%2F4%2F725.atom) 

30. 30.Tuomilehto J, Rastenyte D, Jousilahti P, et al. Diabetes mellitus as a risk factor for death from stroke: prospective study of the middle-aged Finnish population. Stroke 1996;27:210–15 doi:10.1161/01.str.27.2.210 pmid:8571411
    
    [Abstract/FREE Full Text](http://www.ajnr.org/lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6OToic3Ryb2tlYWhhIjtzOjU6InJlc2lkIjtzOjg6IjI3LzIvMjEwIjtzOjQ6ImF0b20iO3M6MTk6Ii9ham5yLzQ2LzQvNzI1LmF0b20iO31zOjg6ImZyYWdtZW50IjtzOjA6IiI7fQ==) 

*   Received August 9, 2024.
*   Accepted after revision September 26, 2024.


*   © 2025 by American Journal of Neuroradiology