Aneta LisowskaAI & Health

Vision, methods & evidence

Research programme

Explainable sequential decision-making across the patient journey, with survival and quality of life as the long-term objectives.

The scientific vision

What sequence of decisions
leads to better patient outcomes?

The goal is measurable: survival and patient-reported quality of life over time. My programme asks how AI introduced at different points in care could contribute to those outcomes, from diagnosis and treatment selection to monitoring, patient support, and recovery.

Reinforcement learning provides a framework for studying policies that optimise expected long-term outcomes. This requires estimating the patient’s current state, modelling possible state transitions given actions, and learning how the consequences of one decision affect the next.

Uncertainty and explainability are integral to this problem: what do the observations tell us, how reliable are predicted consequences, and how can clinicians and patients use that evidence appropriately?

How this programme developed ↗

Three connected research pillars

Understanding, reliance, and action

The pillars serve one programme objective: better survival and quality of life. They connect uncertainty in patient-state estimates, the consequences of actions, and how people use AI in care decisions.

01 · Explainability

What drives a prediction?

Investigating whether explanations help clinicians recognise misleading advice. Clinical meaning, agreement with reader attention, and faithfulness to model behaviour are separate questions.

Explore explainability research ↗
ObserveActLearn

02 · Sequential decision-making

Which action, and when?

Studying policies that optimise expected long-term outcomes. In the oncology vision, the objectives are survival and quality of life; engagement, adherence, response, and toxicity are intermediate measures whose relationship to those outcomes needs testing.

Explore personalized interventions ↗

03 · Uncertainty

How much should we rely on AI?

Estimating patient state from incomplete and variable observations. Uncertainty affects both our assessment at one visit and the changes we infer between visits.

Explore uncertainty research ↗

How does AI influence human decisions?

A decision is an intermediate step. The overarching evaluation asks what AI contributes to survival and quality of life over the patient journey.

Conceptual illustrations of the research themes.

The sequential-decision framework

Patient state. Possible transitions.
Actions with long-term consequences.

AI can contribute at each stage of care by updating what we know about a patient and evaluating what could happen next under different actions.

  1. 01 · Observe

    Gather evidence

    Images, signals, patient reports, and treatment history provide observations across the care journey.

  2. 02 · Estimate

    Infer patient state

    Represent the current state as an uncertainty-aware estimate, rather than assume it is fully observed.

  3. 03 · Predict

    Model transitions

    Estimate how patient state may change given a candidate action, including treatment or supportive intervention.

  4. 04 · Decide

    Evaluate actions

    Study policies that account for consequences over time and support clinically appropriate choices.

Long-term optimisation objective

Maximise survival and quality of life

Evaluate survival and patient-reported quality of life. Test how intermediate measures, including response, toxicity, adherence, and engagement, relate to these outcomes.

Actions are followed by new observations and feedback, informing the next state estimate and decision.

This is the research vision. Current projects investigate parts of the framework; an end-to-end oncology policy and its effect on patient outcomes remain to be developed and validated.

Flagship publications

Contributions to the research programme

Explainability · 2026

Which explanations are clinically impactful?

A systematic review examining clinical impact and evaluation of explainability methods in medical imaging. It connects explanation design with what matters to clinicians.

Shahab Ud Din et al. · Frontiers in Artificial Intelligence · Co-author
↗

Intervention design · 2023

SATO: connecting intervention design with clinical goals

An ontology-aligned workflow for patient-centred mobile behaviour-change interventions spanning multiple clinical goals. My contributions include conceptualization, methodology, analysis, and the original manuscript.

Lisowska, Wilk & Peleg · Journal of Biomedical Informatics · First author
↗

Sequential decisions · 2021

Learning when to deliver a prompt

A feasibility study of reinforcement learning for timely notifications and habit formation on synthetic data. The comparison with adaptive supervised learning exposes the importance of the decision formulation and evaluation baseline.

Lisowska, Wilk & Peleg · SMARTERCARE · First author
↗

Uncertainty · AIME 2026

Recognizing unreliable segmentation

Examines how distributional shift affects uncertainty-aware quality assessment for rare tumor segmentation, testing the reliability of a model intended to assess another model’s output.

Paula Alfken et al. · AIME 2026 · Co-author
↗
Publication record on Google Scholar ↗

Looking ahead · 2027–2031

A research agenda for
AI-supported human decisions

I aim to connect state estimation, action-conditioned transition modelling, and explainable sequential decisions with evaluation of survival and quality of life.

  1. 01

    Explanations that support appropriate reliance

    Compare explanations with radiologists’ gaze and diagnostic reasoning, while separately testing faithfulness to model behaviour. Evaluate effects on diagnostic choices, reading time, confidence, and burden, especially when AI advice is wrong.

    Research question: can an explanation help a clinician detect a mistake, rather than make wrong advice more persuasive?

  2. 02

    Adaptive support under partial observability

    Develop uncertainty-aware patient-state estimates and models of transitions given actions. Study how reinforcement learning can use them to optimise expected longer-term outcomes, including the consequences of intervention timing.

    Research question: how should uncertainty in patient state influence the next action?

  3. 03

    Evaluation across models, people, and outcomes

    Connect technical evaluation and human decision studies with prospective evaluation of survival and patient-reported quality of life. Test whether intermediate improvements translate into the programme’s long-term objectives.

    Research question: when does a better model lead to a better decision?