A Clinical and Imaging Fused Deep Learning Model Matches Expert Clinician Prediction of 90-Day Stroke Outcomes

DISCUSSION

This study demonstrates that a clinical and imaging fused DL model is noninferior to expert physicians in predicting specific mRS outcomes and unfavorable prognoses. Building on our prior work—which established a robust methodology across multiple institutions and demonstrated consistent performance in 2 distinct cohorts—this work not only further enhances the generalizability of the model but also provides critical benchmarks against human expert performance for the task at hand. By evaluating our methodology alongside clinical expert judgments within a unique patient cohort, we investigated the practical implications of our approach in a real-world clinical setting, an element not addressed in our prior work. Slightly better performance was observed in this cohort compared with the prior work, which may be attributed to several factors. These include the following: 1) natural variations in patient demographics and disease presentations across cohorts; and 2) the single-institution cohort of the current study likely offering more uniform treatment protocols, imaging techniques, and patient management strategies, unlike the multi-institution data sets of our prior work, which presented greater variability. Additionally, this work uniquely applies DL to predict stroke outcomes, achieving an ordinal mRS accuracy rate of 36%, nearly triple the rate of random guessing. Prior studies tackling this task typically had lower accuracy using methods such as linear regression and random forest machine learning with hand-crafted features.^1,8

The DLPD also provides practical advantages. By eliminating the need for specialized neurologic expertise, it may be useful for facilities that lack immediate access to neurologists or neuroradiologists. Consequently, it broadens the scope of quality care by making sophisticated prognostic information more accessible across diverse health care environments. The DLPD model uses readily accessible imaging and clinical variables and can be easily integrated into the current clinical workflow for predicting 90-day mRS. This model requires minimal preprocessing steps, with the primary requirement being the normalization of DWI and B0 images to a standard template. Unlike approaches that depend on existing radiologic features—potentially introducing added complexity, human effort, and subjectivity—our method automatically uses information from imaging, thereby lending greater objectivity to our outcome-prediction model. For example, a recent multivariate ordinal mRS regression model required inclusion of 19 separate variables, including some derived from imaging, for which interobserver reproducibility has not been reported.¹⁷ In contrast, the current model relies heavily on objective, data-derived imaging features, with only 7 standard clinical measurements.

This study represents the first report of clinical expert performance for 90-day mRS prediction and allows us to conduct a comparative analysis with the automated model. Thus, the results can act as a benchmark for future studies and further contextualize the results of the DLPD. One prior study¹¹ collected outcome-prediction data from treating providers before endovascular therapy and showed relatively poor performance (44% accuracy to predict into mRS bins of 0–2, 3–4, and 5–6). The performance of the readers in the current study was better, probably because they made their assessments after treatment was provided. Also, these authors stressed that the premorbid mRS was an important predictive feature, but it was often unavailable or inaccurately estimated compared with later retrospective assessment. This issue emphasizes the value of using the entire image, which can incorporate both acute and pre-existing lesions to improve prediction.

Potential reasons behind the noninferior performance of the DLPD in outcome prediction are manifold. First, the DLPD may identify and learn from patterns within complex, multimodal data, similar to or better than how physicians apply their medical knowledge and experience.^18,19 It emulates or improves on human readers by evaluating imaging and clinical data in a data-driven manner, considering all available information, from obvious clinical signs to subtle imaging hints. Additionally, these models can discern and understand nonlinear relationships and interactions among numerous variables, mirroring physicians’ multidimensional thinking when assessing patient conditions and outcomes. Furthermore, DLPD models may bring additional advantages. Their inherent ability to process and learn from vast amounts of data offers unprecedented scalability. With the growth in available data, the performance of the model can potentially increase, indicating a cycle of continuous advancement that may be challenging to match solely with human expertise.

Our study has the following limitations. First, our patient testing cohort was sourced from a single registry. While this feature mimics how the tool might be used in real practice, performance in other cohorts with different characteristics of severity and age is difficult to assess. However, the model has been previously applied to 2 other clinical cohorts with diverse levels of severity and demonstrated similar performance.¹⁴ Second, the mRS served as our primary outcome measure. While the 90-day mRS is widely used to assess chronic disability severity, its subjective determination of categories and variability in reproducibility among different examiners presents notable challenges. Third, the imaging data used in our study were obtained at least 24 hours after the initial baseline imaging; this timeframe was chosen to minimize the effects of any acute interventions, which were completed at the time of imaging. Future studies could consider using initial, pretreatment imaging combined with different therapies to predict outcomes, potentially informing treatment decision-making. Fourth, including imaging sequences beyond DWI and B0 could yield more detailed insight, though this would require more resources and image postprocessing. Fifth, despite the small number of patients in our study, we want to emphasize several factors: Our evaluation was conducted by 5 clinical readers, ensuring a thorough and nuanced assessment, being particularly noteworthy given the complex nature of the reader study and the demanding schedules of the clinicians.

Reader studies are inherently time-intensive, and coordinating such effort among 5 busy clinicians poses substantial challenges. Nonetheless, our findings demonstrate that the performance of the DL model aligns with that of their clinical evaluations, underscoring its potential for clinical application even within a limited patient cohort. This agreement reinforces our hypothesis that the model can operate at a level comparable with that of humans. Last, it is critical to recognize that outcomes may diverge significantly from predictions due to the multifaceted interplay of medical conditions, social determinants, and systemic health care factors that are not entirely predictable by our algorithms. While our model demonstrates robustness, it is not configured to anticipate every acute medical event or the full range of sociodemographic variables that may substantially affect the clinical course. Therefore, there is a clear need to continuously refine predictive methodologies, possibly incorporating a wider set of variables that capture the complexities of patient trajectories; although again, this suggestion comes with drawbacks related to the complexity of the models and the need to collect information that may be difficult to obtain.