Challenges of Outcome Prediction for Acute Stroke Treatment Decisions

Abstract

Physicians often base their decisions to offer acute stroke therapies to patients around the question of whether the patient will benefit from treatment. This has led to a plethora of attempts at accurate outcome prediction for acute ischemic stroke treatment, which have evolved in complexity over the years. In theory, physicians could eventually use such models to make a prediction about the treatment outcome for a given patient by plugging in a combination of demographic, clinical, laboratory, and imaging variables. In this article, we highlight the importance of considering the limits and nuances of outcome prediction models and their applicability in the clinical setting. From the clinical perspective of decision-making about acute treatment, we argue that it is important to consider 4 main questions about a given prediction model: (1) what outcome is being predicted, (2) what patients contributed to the model, (3) what variables are in the model (considering their quantifiability, knowability at the time of decision-making, and modifiability), and (4) what is the intended purpose of the model? We discuss relevant aspects of these questions, accompanied by clinically relevant examples. By acknowledging the limits of outcome prediction for acute stroke therapies, we can incorporate them into our decision-making more meaningfully, critically examining their contents, outcomes, and intentions before heeding their predictions. By rigorously identifying and optimizing modifiable variables in such models, we can be empowered rather than paralyzed by them.

Moving Toward Explainable Decisions of Artificial Intelligence Models for the Prediction of Functional Outcomes of Ischemic Stroke Patients

“…In this section, we focus specifically on studies that use modified Ranking Scale (mRS) ( Table 1), (16) , as the main functional outcome measure, because it is a common form of communication across all those involved in the patient's care pathway (7), has good reproducibility (16), is the most prevalent functional outcome measure in contemporary stroke trials (17) and is therefore frequently used as the primary outcome measure in AI prediction studies (3,4) . However, it is worth noting that the discussion points of this chapter are relevant to most AI-based prediction studies in stroke.…”

“…In their current form, even the most advanced models do not perform well enough to be implemented in the clinical setting (4). There are several limitations in functional outcome prediction that contribute to these models' inability to go beyond a prediction accuracy ceiling.…”

“…CDSSs will be particularly beneficial in the care of ischemic stroke patients due to the heterogeneity, complexity, and time-critical nature of the condition and the wealth of physiological information available from neuroimaging. AI algorithms can provide decision support for ischemic stroke in various clinical tasks, such as diagnosis, onset time estimation, and prognosis (3)(4)(5). However, AI systems can be met with resistance from health care professionals because even if the system is accurate, it is not obvious how it arrives at its decisions; this is the so-called ''black box problem'' (6).…”

Machine Learning Models for Predicting Influential Factors of Early Outcomes in Acute Ischemic Stroke: Registry-Based Study

“…Ischemic stroke survivors commonly have disabilities and substantial function loss that significantly affect their quality of life. Outcome prediction provides a reference for doctors to select rehabilitation strategies and provides patients with decent expectations in the future [ 2 , 3 ]. Several studies have focused on stroke prediction by indicators collected at emergency room (ER) or first at ward admissions [ 4 , 5 ].…”

Full Reperfusion Without Functional Independence After Mechanical Thrombectomy in the Anterior Circulation

“…With an ever-growing number of patients being treated with MT worldwide over the last years, a broad range of factors predicting good clinical outcome after MT were described. Based on these, a plethora of outcome prediction models were developed with different purposes and variable settings [ 2 ]. The majority of prediction models focus on good clinical outcome after treatment measured by the modified Rankin Scale (mRS), which consists of seven levels between no clinical deficit (mRS = 0) and death (mRS = 6) [ 3 ].…”