Editorial: Big Data analytics to advance stroke and cerebrovascular disease: a tool to bridge translational and clinical research
Year: 2023
Type: editorial
Abstract: Big data analysis has the potential to enhance the high through put processing required to better phenotype patient outcomes post treatment, select potential therapeutic targets, and refine biomarker selection for risk assessment and disease monitoring. (1) With data registries, more advanced imaging, data storage tools, and more robust electronic clinical documentation, robust analysis can be conducted with large datasets with very granular individual patient level data. (1,2,3) Analysis of large datasets requires special considerations to ensure that the significant associations or findings are clinically meaningful and without bias. (1) Use a Big Data approach can aid in the discovery of pertinent biomarkers for diagnosis and assessment of stroke risk. Wu et. al, used regression modeling to determine which factors were associated with patients have brain infarction detected on MRI in a cohort of 1.4 million patients living in China, demonstrating that there were geographic, sex-related, and metabolic disease risk factors for having infarction detected on brain MRI.(4) Efficacy of anticoagulant type was compared by Lee et. al., demonstrating a lower risk of stroke and bleeding complications associated with nonoral vitamin K antagonists.( 5) Liao conducted a study including over 5 million patients to confirm the increased risk of stroke in association with markers of insulin resistance.( 6) Shu et. al demonstrated that altitude has an increased risk of the development of ischemic changes on MRI and an inverse relationship with risk of clinical events of acute stroke. (7) Yeng et. al studied the efficacy of several machine learning models to predict genetic stroke risk include the LASSO, artificial neural network, and random forest, and support vector machine-recursive feature elimination model, showing that there are limitations using this approach as their models were limited in their accuracy and specificity. (8) Another approach that can be useful are Mendelian randomization models. Ma et. al. were able to demonstrate that genetic variants previously demonstrated to be associated with elevated homocysteine levels were not associated with an increased risk of intracranial aneurysm detection by using several Mendelian randomization models. (9) Hao Wang were able to use random forest models to better predict risk of subarachnoid hemorrhage in patients with MCA aneurysms. (10) Combining imaging and clinical variables can improve patient phenotyping. Yingwei Guo et. al. investigated machine learning models as a diagnostic tool to diagnosis stroke by automation. (11) For example, Yi Li demonstrated that CT imaging features and markers of small vessel disease are predictive of presence of the presence of > 10 cerebral microbleeds on MRI. (12) More research is needed before big data analysis such as artificial intelligence and machine learning requires can be more ubiquitously applied to clinical care. (2,3,13,14,15) Having practical models that allow for quick assessment of risk for hemorrhagic conversion and risk factors for hemorrhagic conversion have the potential to help with stroke stratifying risk of revascularization therapies such as thrombolysis as special considerations have to be considered based on clinical factors such as duration of symptoms, (16,17) medications, imaging, and clinical comorbidities within 4.5 hour window and in the extended time window per the American Heart Association Guidelines on acute ischemic stroke management.(17) Ren et. al. used modeling and area under the curve receiver operation characteristic curve analysis to develop a score for predicting risk of hemorrhagic conversion with thrombolysis with an area under the curve value of 0.82.(18) Miaomiao Yang et. al. created a nomogram that predicts stroke risk with thrombolysis using a combination of imaging, clinical, and blood biomarkers. (19) Risk of ischemic hemorrhagic conversion associated with thrombolysis is further reviewed by Huiling Shao et. al. (20) Machine learning can also be used to parse areas of cerebral hypoperfusion and areas of normal cerebral perfusion, which is information that has been useful in thrombectomy clinical trials and in early patient selection tools for thrombectomy. Machine learning has been investigated for its diagnostic utility. Xinping Li demonstrated that early patient characteristics available within the first 24 hours of hospital admission can be predicted of early outcomes post thrombectomy.( 21) They were able to fine tune those predictions using different models such as a the SHapley Additive exPlanations approach. (21) Modeling can also be useful in investigations on posterior circulation infarction such as basilar artery occlusion. Chenhao Zhao et. al. confirmed that risk factors such as atrial fibrillation increase risk of recurrent stroke but do not influence stroke outcomes basilar artery thrombectomy outcomes.( 22) While, Lin et. al developed a nomogram to help predict in which patient's with basilar artery occlusion recanalization would be futile.(23) Zeng et. al., also looked at futility, but focused on thrombectomy outcomes in the anterior circulation and using a combination of machine learning models with combined with using the stacking method.(24) Currently, the American Heart Association only endorses volumetric analysis for thrombectomy patients in the extended 24 hour window.(17) However, several large core endovascular trials have subsequently demonstrated that even patients with large cores may still have some benefit from thrombectomy. (25,26,27) More research is needed to optimize prediction tools for patient selection for thrombolysis and thrombectomy.Cost of stroke care is projected to be over $90 billion dollars by 2035. Part of those costs are attributed to extra healthcare costs related to stroke associated morbidity. (1,28) Determine who is at risk of medical complications associated after a stroke and tailoring a post stroke recovery plan could be quite impactful.(1) Ji et. al. used modeling to develop a risk score to predict the risk of a patients admitted with intracerebral hemorrhage developing an deep vein thrombosis while they are hospitalized optimized with external validation cohort.( 29) With studies, comparison of machine learning models can demonstrate which model provides the best sensitivity and specificity to predict the clinical outcome of interest. For example, Zheng et. al. compared several machine learning models to determine which model would be most specific and sensitive for predicting which patients admitted with an intracerebral hemorrhage would have post stroke course complicated by the development of pneumonia, showing that the Gaussian naïve Bayes and logistic regression models both performed well depending on whether the internal or external validation cohorts were used.(30) Shuo Feng et. al. demonstrated similar proteins elevated during thrombotic events (acute myocardial infarction and acute ischemic stroke), identifying markers of inflammation. (31) In a study including over 100,000 intracerebral hemorrhage patients, Zhao et. al. combined regression analysis and with causal mediation analysis to determine driving factors behind sex-related outcomes, showing the hemorrhage location and clinical severity were the strongest driving factors of mortality and morbidity.(32) Gu et. al demonstrated that mortality rates are higher in critical ill patients with intracerebral hemorrhage and low calcium levels. (33) Others have used Big Data analytic approaches to study length of stay, healthcare utilization, healthcare costs.(3) Currently, there are no widely accepted models for predicting morbidity and mortality for clinical purposes.Big Data analysis has been studied to provide prediction models to improve management and coordination of post-acute care. Resource utilization post stroke and needs can vary in patients after hospital discharge and best practices for managing stroke recovery can change over time. (34) Prediction models can be used to determine which risk patient characteristics are the most associated with likelihood of hospital re-admission within 30 days, (35) which can used to better allocate resources and services for patients. Yu-Ching Chen et. al. compared multiple models and assessed the sensitivity and specificity of machine learning models to select the best machine learning model predicted readmission within 30 days of hospital discharge.(35) Yarfi et. al. propose using mixed methods models and qualitative analysis to assess post stroke rehabilitation outcomes.(36) Boutros et. al. demonstrated that depression was associated with recurrent stroke and mortality one year after stroke. (37) Another model that can be useful in analyzing clinical trial data is Bayesian Network Meta-analysis. Li et. al. evaluated several approaches for addressing post stroke cognitive dysfunction, and found that transmagnetic stimulation and acupuncture could be helpful.(38) Chen et.al. demonstrated that machine learning can be used to differentiate responses to transcranial magnetic stimulation between patient's during the post stroke recovery phase by using unsupervised hierarchical clustering, which could have utility in tracking post stroke recovery. (39) In addition, several studies have used a Big Data approach for assessing quality of life indices. (3) Big data analytics is a rapidly evolving field and there are important considerations and pauses that should be factored into data interpretation and application. It is important to be aware of biases that by be present in datasets as a result of patient recruitment. (1,2,3,13,14,15) Even within large datasets, there may be unknown missing confounders. It is important to consider validation of results in different datasets. (1,2,3,13,14,15)
Source: Frontiers in Neurology
Language: English
Subfield: Epidemiology
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $3,295
Funder: National Institute on Aging
Award: G1653274783
OpenAlex ID: W4389923920