Prescriptive AI tools shift critical care toward personalized treatment.

Critical care requires rapid decisions from complex, evolving data. Predictive models forecast deterioration but do not guide therapy. Prescriptive AI-causal machine learning (ML) and reinforcement learning (RL)-can bridge this gap. Causal ML estimates individualized treatment effects; RL generates adaptive treatment policies over time. This review examines the shift from prediction to personalization, key barriers, and future pathways including digital twins and clinician‑in‑the‑loop systems.

Introduction

Critical care demands rapid interventions informed by evolving data. Decisions on fluids, vasopressors, or PEEP determine recovery. Each patient has a unique trajectory, and ICUs generate heterogeneous data-labs, vitals, notes, waveforms, imaging. Personalization requires individualized estimates anticipating the consequences of clinical actions, not population averages. AI provides the foundation: predictive models forecast deterioration, ML estimates intervention effects, and RL optimizes decision sequences.

Current state of AI

Most AI applications remain predictive. Early warning systems forecast deterioration; models predict AKI, sepsis, or mortality. AI has identified subphenotypes of sepsis, AKI, and ARDS, revealing biologically distinct clusters. Multimodal models now integrate records, waveforms, imaging, and free‑text notes using recurrent networks, transformers, and large language models. However, these systems estimate likelihoods but do not recommend interventions. A model may forecast mechanical ventilation but not advise whether to escalate support, adjust sedation, or initiate prone positioning. This gap points to the need for causality‑aware, sequential decision models.

Achieving true personalization

Causal ML: Causal ML estimates heterogeneous intervention effects rather than associations. It defines causal estimands-average treatment effect, individualized treatment effect, and the value of an individualized treatment rule. In critical care, treatments are rarely randomised, introducing confounding. Credibility relies on conditional exchangeability, positivity, consistency, and correct model specification. Methods include propensity matching, inverse probability weighting, causal forests, meta‑learners, and longitudinal g‑methods. These can answer whether fluids help or harm a patient in septic shock, or whether higher PEEP improves oxygenation without worsening haemodynamics. Classical models forecast AKI but cannot tell whether giving fluids improves outcomes. A recent study used causal forests and a policy tree to identify septic patients with AKI likely to benefit from restrictive fluids; those predicted to benefit had higher AKI reversal rates.

RL: RL optimizes sequences of interventions over time. Critical care is dynamic: insulin titrated hourly, ventilator settings adjusted, fluid/vasopressor balance revisited. RL formalises this as a Markov decision process with states, actions, transitions, and rewards, aiming for a policy maximising cumulative reward. Because trial‑and‑error on critically ill patients is impossible, applications use offline RL, learning from retrospective data under historical practice. Careful off‑policy evaluation-importance sampling, doubly robust methods, fitted Q evaluation-is essential. RL has been applied to fluids and vasopressors in sepsis, ventilator management, and glycemic control after cardiac surgery, with validation showing recommendations at least as safe as those of experienced clinicians. Prospective trials are needed.

Challenges

Data quality: Records contain missingness, delayed documentation, unit inconsistencies, and device artifacts. Implementation requires robust data engineering, real‑time ingestion, and harmonisation. Lack of interoperability impedes integration.

Interpretability: Clinicians need to know which features drive a recommendation. Explainable AI and human‑centred design remain under‑explored in high‑acuity settings.

Workflow integration: Systems must fit fast‑paced, team‑based workflows. Poorly timed alerts add burden. Co‑design and usability testing are essential. Clinicians must retain authority, with override mechanisms logged. Thresholds should be tuned to minimise alert fatigue.

Privacy and security: Real‑time interfaces require authentication, role‑based access, and audit logging.

Prospective validation: Retrospective accuracy cannot validate RL and causal ML. Off‑policy evaluation approximates performance but cannot anticipate dynamic environments. Silent deployment and simulation offer lower‑risk pathways. The field lacks standardised evaluation frameworks.

Documentation and governance: Model cards, data provenance, and continuous drift monitoring are needed. RL policies may evolve, complicating oversight. Frameworks such as Good Machine Learning Practice and SaMD guidance emphasise intended use, risk management, and change control.

Fairness: Historical data reflect inequities. Continuous auditing and fairness‑aware retraining are required.

Future directions

Causal ML should move toward reliable treatment effect estimates using target trial emulation, time‑varying confounding methods, and high‑dimensional g‑methods. Studies should prespecify estimands, report positivity diagnostics, and include external validation. RL will remain offline; conservative Q‑learning, distributional RL, and behaviour‑regularised methods limit unsafe extrapolation. Multiobjective rewards balancing competing priorities bring policies closer to clinical trade‑offs. Prospective validation should follow a path of retrospective development, off‑policy evaluation, simulation, silent deployment, and clinician‑in‑the‑loop pilots. Digital twins-virtual patients blending physiology with real‑time data-could provide safe testing environments. Reporting should follow TRIPOD‑AI, PROBAST‑AI, DECIDE‑AI, SPIRIT‑AI, and CONSORT‑AI.

Limitations

This is a narrative review; study selection reflects editorial judgment. The literature remains largely retrospective and proof‑of‑concept, with no prospective trials showing improved outcomes.

Conclusions

AI can move critical care beyond prediction toward individualized prescription. Causal ML estimates treatment effects; RL optimizes sequences of actions. Challenges remain-data quality, interpretability, workflow integration, validation, governance, and equity-but digital twins offer pathways for safe testing. These systems are decision‑support tools; clinicians retain ultimate authority. If responsibly developed, they may enable the right treatment, for the right patient, at the right time.