Source-linked AI summary
Artificial Intelligence Should Genuinely Support Clinical Reasoning and Decision Making To Bridge the Translational Gap
Kacper Sokol, James Fackler, Julia E Vogt
TL;DR
Medical AI has struggled to cross the translational gap because technology-centric systems can misalign with clinical reasoning, workflows and institutional constraints. This Perspective proposes sociotechnical AI that augments clinicians’ cognitive and epistemic functions, concluding that human-aligned support is a promising route toward real-world impact.
Problem
Medical AI faces a translational gap because systems can be incompatible with clinical reasoning, decision making, workflows and broader sociotechnical requirements.
Method
The Perspective proposes integrating AI into clinical systems ecology as a human-centred supporting tool informed by cognitive sciences and aligned with real-world decision protocols.
Results
The Perspective concludes that positioning AI as a partner or tool that complements clinicians’ expertise and cognitive functions is a promising way to address medicine’s translational barrier.
Takeaways & Limitations
AI should prioritise supporting perception, reasoning and decision making over pursuing superhuman performance on narrowly defined benchmarks.
Takeaways & Limitations
Semi-structured clinical tasks may resist end-to-end modelling because complex healthcare decisions may lack one optimal solution and require subjective judgement.
Abstract
from arXiv · showhide
Artificial intelligence promises to revolutionise medicine, yet its impact remains limited because of the pervasive translational gap. We posit that the prevailing technology-centric approaches underpin this challenge, rendering such systems fundamentally incompatible with clinical practice, specifically diagnostic reasoning and decision making. Instead, we propose a novel sociotechnical conceptualisation of data-driven support tools designed to complement doctors' cognitive and epistemic activities. Crucially, it prioritises real-world impact over superhuman performance on inconsequential benchmarks.
Introduction
AI’s clinical impact remains constrained by a translational barrier: technology-centric systems often fail to align with clinical practice and decision making. The Perspective proposes human-centred tools that support doctors’ cognitive and epistemic functions rather than replace them.
- AI’s healthcare success stories remain scarce relative to the number of systems being developed, despite potential benefits for access, efficiency, prevention and treatment.
- The translational barrier reflects technical misalignment, workflow friction, limited user acceptance and unresolved fairness, accountability, robustness and interpretability concerns.
- The proposed sociotechnical approach aligns AI operation with doctors’ needs, expectations and clinical environments through an interdisciplinary, human-centred view of intelligence.
- Cognitive-science-informed tools could support reasoning under noise and uncertainty, improve decision consistency, reduce reasoning errors and make thought processes more principled.
- Paediatric sepsis exemplifies urgent clinical reasoning challenges because its incidence, treatment strategy, diagnostic criteria and predictive indicators remain uncertain.
Medical Artificial Intelligence Adoption Challenges
Medical AI adoption is hindered when generic predictive models ignore temporal trajectories, clinical workflows, institutional context and sociotechnical constraints. Paediatric sepsis exposes these shortcomings through ambiguous diagnosis, heterogeneous patients and uncertain progression and treatment.
- Paediatric sepsis combines ambiguous recognition, uncertain risk and progression assessment, inconsistent treatment, limited response monitoring and age-group heterogeneity.
- Conventional point-in-time predictions can assign identical outputs to patients with the same current state despite diverging trajectories toward recovery or critical care.
- AI systems that model temporality are more appropriate for trajectory-based care, yet remain broadly underutilised.They could address questions about antibiotic timing and future critical-care needs.
- Healthcare AI research often targets narrow benchmark tasks selected for data availability rather than bespoke clinical applications.
- Clinicians may reject ill-conceived systems when they impose cognitive burden, disrupt workflows, fail to provide needed information or conflict with practical constraints.
- Medical AI must also address bias, representativeness, fairness, privacy, interpretability, reliability, robustness, accountability, regulation, workflow compatibility and resource constraints.
Systems Ecology and Artificial Intelligence
The paper argues that clinical AI should be designed as sociotechnical support for doctors’ reasoning and decision making, rather than as autonomous systems delivering ready-made conclusions. Such tools must fit clinical workflows, organisational structures, and human cognitive and epistemic needs.
- Misalignment with clinical work: Current decision-support tools often present conclusions rather than genuinely supporting decision making, competing with and curtailing clinicians’ judgement.This misalignment can bias perception, impede cognition, limit independent reasoning, and erode expertise.
- Sociotechnical design: AI development should begin with concrete decision makers’ needs and account for the workflows, constraints, and broader environments in which decisions occur.The proposed systems would augment comprehension, exploration, problem solving, and prospective critical reasoning.
- Accountability: Full automation can create ethical problems because responsibility becomes unclear when algorithmic decisions cause unintended harm.Augmenting human decision making keeps responsibility with humans within established frameworks such as evidence-based medicine.
- Distributed cognition: Human–machine symbiosis and distributed cognition can allocate responsibilities across specialised human and algorithmic agents while preserving meaningful human decision making.Different activities may be automated, assisted, or augmented according to their role and context.
- Organisational embedding: Clinical tools can function as safely usable black boxes when organisational mechanisms assign responsibility for reliability, certification, calibration, and operation.Laboratory tests and MRI illustrate how institutional structures make complex technologies usable in routine care.
- Human–AI partnership: AI is more acceptable as a digital partner that complements human abilities and decision making than as a replacement for human intelligence.Support may add evidence, compensate for weaknesses, prevent biases, and overcome limitations without reducing clinicians to accepting or rejecting recommendations.
Human Decision Making and Artificial Intelligence
The paper distinguishes AI uses according to human–AI roles, task structure, and environmental stability. Because clinical diagnosis often involves uncertainty, subjective judgement, and changing conditions, support for human reasoning may be preferable to end-to-end automation.
- AI deployment modes: AI deployment can range from autonomy and human-in-the-loop assistance to machine-in-the-loop augmentation of higher-level human cognition.The most suitable paradigm should be selected separately for activities such as data acquisition, analysis, decision selection, and implementation.
- Decision environments: Automation is most suitable where environments provide reliable regularities, whereas wicked environments lack dependable cues, feedback, or predictability for complete automation.Naturalistic Decision Making studies successful expert judgement, while Heuristics and Biases examines faults in basic reasoning.
- Clinical task structure: Clinical diagnostic reasoning is often semi-structured because incomplete information, uncertainty, subjective judgement, and unpredictable outcomes prevent one universally optimal solution.Rigid algorithmic recommendations may suppress exploration of alternatives and impede the development of new knowledge.
- Model choice: In open-world tasks, simple transparent models or high-level heuristics can perform as well as or better than complex data-driven systems.The paper contrasts this with the stronger performance of large datasets and advanced algorithms in stable-world structured settings.
- Paediatric sepsis: Neonatal sepsis illustrates both expert performance that can be studied and a wicked environment in which uncertainty produces variable decisions.Naturalistic Decision Making helped identify infection indicators that were validated across hospitals and formalised into staff instruction.
- Cognitive support: Cognitive biases such as time preference can favour immediate antibiotic treatment over future benefits such as preventing antimicrobial resistance.The paper therefore favours supporting doctors’ cognitive and epistemic functions over merely modelling and improving their observed actions.
From Benchmark to Bedside
The paper moves from benchmark-oriented prediction toward AI tools that strengthen clinicians’ reasoning in complex, uncertain environments. It proposes cognitive support such as explanation, hypothesis generation, mental simulation, and multi-scenario planning, including applications for training and clinician comparison.
- Cognitive support: AI tools could strengthen doctors’ knowledge, critical thinking, and reasoning while preventing biases and improving decision hygiene.Existing explanation research can mitigate some reasoning faults but may exacerbate others and often overlooks the wider systems ecology.
- AI toolkit: Potential capabilities include extracting human-comprehensible concepts, testing hypotheses, supporting mental simulation, and prompting analogical or counterexample-based reasoning.These functions are intended to support clinicians’ reasoning rather than replace it with a prediction.
- Mental simulation: Simulation tools could let doctors project patient trajectories, test hypothetical scenarios, and compare pathways while retaining missing or unknown information.Alternative trajectories can be conditioned on different values of a selected variable, such as a medical test result.
- Robust decision making: Planning across multiple probable outcomes can improve robustness by making critical junctures, complications, and unexpected events explicit.The proposed support may reduce cognitive fatigue, overconfidence, and errors while improving decision consistency in stressful clinical settings.
- Training and collaboration: Simulation can also provide safe environments for mental rehearsal, case-based training, and learning from landmark clinical decisions.The paper additionally proposes comparing digital twins of clinicians or using them as benchmarks for junior doctors.
Conclusion
The Perspective proposes sociotechnically integrated AI that supports clinicians’ cognitive and epistemic functions rather than replacing human judgement. It argues that aligning tools with clinical workflows and preserving human responsibility could improve acceptability, decision making, and outcomes such as paediatric sepsis detection and management.
- The Perspective shifts AI operationalisation from superhuman performance toward supporting human perception, reasoning, and decision making.
- Its sociotechnical approach integrates AI with clinical workflows and decision-making protocols while preserving human responsibility and autonomy.
- The approach aims to minimise decision errors and undesired variability while making judgement more factual, evidence-based, and principled.
- The proposed tools provide timely, relevant insights that complement clinical expertise, support scenario reasoning under uncertainty, and help clinicians reach more robust diagnoses.
- In paediatric sepsis, AI could aid detection, improve severity assessment and progression monitoring, and make treatment more consistent to reduce unnecessary antibiotic exposure.
- Integrating AI into multidisciplinary sepsis teams and bedside huddles could support participants’ cognitive and epistemic functions.