From Average Treatment Effects to Individualized Decision Making in Acute Ischemic Stroke
September 4, 2026
Mohamed F. Doheim, MD, PhD, FRCP, is clinical research assistant professor, Department of Neurology, University of Pittsburgh School of Medicine. Dr. Doheim is the founding director of the NeuroPrecision Lab at the UPMC Stroke Institute. He earned his medical degree from Alexandria University in Egypt and completed his PhD at Maastricht University, where his work focused on precision medicine in acute ischemic stroke.
Dr. Doheim also completed a National Institutes of Health StrokeNet Fellowship at the UPMC Stroke Institute and is currently pursuing a Master of Data Science at the University of Pittsburgh. Dr. Doheim’s research examines how clinical trial evidence, large clinical datasets, causal inference, machine learning, and decision-support systems can more effectively personalize treatment for strokes. His research involves developing individualized treatment selection, thrombectomy technique, stroke triage and transfer, neuroprotection, biomarker discovery, clinical trials, and device development.
In the following interview with Dr. Doheim, he discusses the problems that organize his research, the distinction between technical and clinical success in stroke care, the limitations of unaided decision-making in a time-sensitive setting, and the novel tools he and colleagues at the UPMC Stroke Institute are developing to estimate how individual patients experiencing an acute stroke may respond to different treatment strategies.
Q: Why is acute ischemic stroke poorly suited to a one-size-fits-all treatment model?
A: Simply stated, acute ischemic stroke is not a single disease entity, and average treatment effects do not translate uniformly to every patient. A stroke can be minor, moderate, or severe. It can involve a large vessel or a small distal vessel. It can arise from atrial fibrillation, atherosclerosis, or another mechanism. The occlusion can occur in different regions, collateral circulation can vary, the clot itself can have different characteristics, and the patient may arrive at a different time with a different level of baseline function and different comorbidities.
I sometimes compare stroke with a symphony. You have multiple instruments, and the way they are playing together produces what you hear. Stroke has many components interacting at the same time. Two patients may carry the same diagnosis but have different vascular anatomy, brain resilience, clot composition, and probabilities of responding to treatment.
Q: Clinicians already consider age, stroke severity, imaging, and time when determining the best treatment approach. What can a computational model add that cannot be done by individuals or teams alone?
A: The problem is the number of variables, the interactions among them, and the speed at which the treatment decision has to be made. The patient may move from emergency medical services to a primary stroke center, through telestroke evaluation, and then to a comprehensive stroke center. At each point, clinicians are making treatment and transfer decisions with limited time.
As human beings, we are good at linear processing, but we are not as good at nonlinear processing. If you tell me that a patient is 85 years old, I understand that information. If you ask me to combine age with stroke severity, occlusion location, time, collateral status, glucose, prestroke disability, and the interactions among them, I cannot calculate that consistently in my head within a minute or two.
We also face a form of anchoring bias that we sometimes call the "eyeball fallacy.". One characteristic attracts our attention, and we anchor on it. A clinician may focus on age and underweight the patient’s functional status, the vessel involved, or another characteristic that changes the probability of benefit.
Experience remains important, but an individual clinician’s experience is still based on the patients that person has seen. A model can compare the current case with hundreds or thousands of prior patients and weight the variables consistently.
The purpose is not to replace clinical judgment or shared decision-making. It is to provide an objective estimate that the physician can incorporate with the examination, imaging, technical feasibility, patient preferences, and clinical judgment.
Q: How is individualized treatment estimation different from ordinary prognostication?
A: Traditional prediction often occurs after treatment has occurred. We can look at the patient’s neurologic status after thrombectomy, for example, and estimate the likelihood of a favorable outcome at 90 days. That information can help with counseling and planning, but it cannot change the treatment decision that has already been made.
Individualized treatment-effect estimation asks a different question: how might the same patient's outcome differ under alternative treatment strategies? Before treatment, the model estimates what may happen if this patient receives endovascular therapy and what may happen if the same patient receives medical management. The difference between those probabilities is the estimated treatment effect for that patient.
One way to conceptualize this patient-level counterfactual model is as a "digital twin," although the estimate remains probabilistic rather than deterministic. With enough data, we can identify prior patients who resemble the current patient and examine how they responded under different treatments and levels of care. We are not claiming to know with certainty what will happen, and confidence in the estimate depends on the quality, completeness, and representativeness of the underlying data. We are using those experiences to provide a comparative estimate before the clinician has to choose a path.
The model also has to be interpretable. The clinician should be able to see the variables used, the probability estimated under each treatment, the magnitude of the difference, and the limitations of the underlying data. .A score without that context, including an explicit representation of uncertainty, is not enough.
Q: Why is the distinction between technical success and clinical success in stroke care important?
A: With thrombectomy, the field currently achieves successful angiographic reperfusion in approximately 85% to 90% of cases. We can reopen the occluded artery and restore macrovascular blood flow Functional independence, however, remains below 50%, and a substantial proportion of patients remain dependent or die despite technically successful reperfusion. A procedure can be technically successful without enabling the patient to walk again, return to independence, or even survive. We call that the recanalization paradox.
Many factors go into this. The brain may already have sustained substantial injury before reperfusion. The patient may have limited collateral circulation or baseline frailty. The procedure and anesthesia can affect hemodynamic stability. Small clots may continue to obstruct the microcirculation even when the artery appears open on angiography.
Reperfusion itself can also contribute to injury. I describe it this way: You open the road, but there is still a fire in the neighborhood. Opening the road allows blood flow to return, but tissue recovery still depends on microvascular integrity, collateral support, the extent of established injury, and the biologic response to reperfusion. . That gap between recanalization and functional recovery is one reason we are also studying neuroprotective approaches and biomarkers that may help identify additional therapeutic targets.
Q: How are you and the UPMC Stroke Institute applying these approaches to distal and medium vessel occlusion stroke?
A: Distal and medium vessel occlusions make up about 25% to 40% of ischemic stroke cases, but we do not yet have a universal treatment strategy for these patients. The vessels are smaller and more distal. Deficits may appear less severe on an aggregate score but can still be quite disabling. The technical risk of thrombectomy and the potential benefit both vary considerably.
The multicenter Distal and Medium Vessel Occlusion Stroke (DUSK) trial our center created and led evaluated endovascular treatment in this population. Separately, we developed the DUSK Tool to estimate the probability of functional independence under endovascular therapy and under medical management for an individual patient.
The tool uses information collected in routine care, including demographic characteristics, clinical assessment, laboratory and imaging findings, time, intravenous thrombolysis treatment, and medical history. We deliberately avoided making it dependent on advanced variables that may be unavailable in a community hospital. If a decision-support tool is intended for a regional system, it must be pragmatic.
The output might show, for example, a 60% estimated probability of functional independence with endovascular therapy and 53% with medical management. The seven-percentage-point difference represents the estimated individualized treatment effect and gives the clinician a measure of the potential incremental benefit. In another patient, medical management may produce the better estimate.
The DUSK Tool is currently for research and education. It requires prospective validation before it can be deployed in routine care. The broader idea is that a neutral population-level result may conceal meaningful treatment-effect heterogeneity: some patients may benefit substantially, others minimally, and some may be harmed. A neutral average result should therefore prompt closer examination of individual responses rather than the assumption that every patient has the same expected treatment effect.
Q: Beyond the decision whether to perform thrombectomy, what other treatment choices may benefit from a more individualized approach?
A: Anesthesia is one example. A patient who is agitated, anxious, or unable to remain still may need general anesthesia for the procedure to be performed safely. Another patient may tolerate monitored anesthesia care or local anesthesia. General anesthesia can produce good technical results, but hypotension and other hemodynamic instability during the procedure may reduce the likelihood that technical success becomes clinical success.
We have also studied whether intravenous thrombolysis should be given before thrombectomy. The answer may differ depending on whether the patient presents directly to a thrombectomy-capable center or is transferred from a primary stroke center. Occlusion location, timing, and other characteristics may modify the effect of combining treatment options.
These are examples of the same problem. The question is not simply whether a treatment or technique works. It is which patient is most likely to benefit from it, under which circumstances, and at what risk.
Q: How does the NeuroPrecision Lab connect computational research with clinical and translational work?
A: The lab includes machine learning, personalized care, causal and statistical analysis, and clinical trials, but those are only part of the program. We are also developing tools that can integrate evidence into the clinical workflow.
Several ongoing projects focus on translating evidence into clinical workflows. LUCID is designed to tailor relevant guideline recommendations to an individual patient in real time, while Trial Lens is intended to facilitate matching with appropriate clinical trials when treatment uncertainty remains. We are also working on devices, biomarkers, and neuroprotective strategies. These projects are connected by the same objective: identify the right treatment, improve how it is delivered, and develop another option when the existing treatment is insufficient. Data analysis can reveal the problem, but translation may require a clinical trial, decision-support system, biomarker, drug strategy, or device.
Q: What must happen before these kinds of decision-support systems become part of routine stroke care?
A: Model performance is only one requirement. The system has to use information available early enough to affect the decision. It must work across different hospital environments, integrate with the clinical workflow, and communicate the estimate in a way that is understandable within seconds. It also has to undergo prospective validation, security review, governance review, and continued evaluation after deployment.
The evidence changes quickly. A treatment recommendation that was reasonable last year may no longer reflect current trials or guidelines. A decision-support system can help bring current evidence to the point of care, but it has to be updated and monitored carefully.
We also have to study whether use of the tool changes decisions and improves outcomes. A model may predict accurately in a retrospective or external dataset and still fail to help in clinical practice if it arrives too late, interrupts workflow, or is not trusted by clinicians. The goal is not merely to build an accurate model. The endpoint is whether the patient receives the treatment most likely to produce the best functional outcome.
Q: Why can the term “minor stroke” be misleading?
A: A low NIHSS score does not mean the stroke is benign. Consider a surgeon who loses a very fine hand movement or a pianist who loses the movement needed to perform. The numerical deficit may be small, but the effect on that person’s career and life can be profound. It matters who the patient is, what the patient does, and which function has been lost.
Our work on thrombolysis in minor stroke distinguishes disabling from nondisabling deficits. Across patients with nondisabling symptoms, thrombolysis may not provide sufficient benefit to justify the risk. A disabling deficit may warrant different consideration after assessment of the individual risks and expected benefit.
We have proposed a framework that looks beyond the aggregate severity score in stroke. We consider the vessel involved and the functional territory it supplies, the collateral circulation and resilience of the brain, and the characteristics of the clot. The word “minor” should not end the analysis. Ultimately, precision stroke care is not only about predicting outcomes more accurately; it is about making better treatment decisions for the individual patient in front of us.
Learn More
