Personalized Fluid Management in Patients With Sepsis and Acute Kidney Injury

• Causal‑ML pipeline (causal forest → policy tree) flagged ICU patients with sepsis‑associated AKI who may benefit from restricting IV fluids to ≤500 mL in the 24 hours after AKI onset; in external validation, flagged patients who received restriction had early AKI reversal 53.9% vs 33.2% and MAKE30 (major adverse kidney events at 30 days) 17.1% vs 34.6%.

• Model prioritized heterogeneous benefit better than a noncausal random forest (development AUTOC 0.73 vs 0.45; external AUTOC 0.15 vs −0.02), but real‑world uptake was low (15.6% development; 5.7% external), so treat as a triage signal, not an automated stop.

• Post‑resuscitation IV‑fluid practice after AKI onset is variable and can increase dialysis risk and ICU resource burden; pharmacy typically owns IV‑fluid stewardship, compounding, and verification workflows.

• Untargeted alerts or poor data feeds will add work, erode clinician trust, and limit adoption; a high‑precision triage tool could concentrate pharmacist effort where it matters.

• Retrospective EHR study using MIMIC‑IV (development n=11,650) and SICdb (external n=1,931) of adults with Sepsis‑3 within 24 h of ICU admission and KDIGO‑defined AKI within 48 h.

• Exposure: cumulative IV fluids from AKI onset → 24 h; restrictive defined in text as ≤500 mL (some excerpts show <500 mL).

• Causal forests estimated individualized treatment effects; a policy tree translated ITEs into interpretable bedside rules using routine vitals, labs, weight, and urine output.

• No prospective deployment; evaluation via AUTOC and outcome comparisons within model‑recommended groups.

• External recommended‑patient results: among patients the model recommended for restriction, adjusted odds in external validation were early AKI reversal OR 2.22 (95% CI, 1.37–3.60), sustained reversal OR 2.07 (95% CI, 1.24–3.45), and MAKE30 OR 0.45 (95% CI, 0.23–0.86).

• Performance gap: AUTOC fell from 0.73 (development) to 0.15 (external), indicating attenuated out‑of‑sample policy value; a conventional random forest performed worse at both sites.

• Adoption gap and missing operational detail: despite predicted benefit, very few patients received ≤500 mL (5.7% external); exact policy‑tree thresholds, node sizes, baseline‑creatinine rules, and raw subgroup counts were not reported in the excerpts.

• Context note: large RCTs of restrictive vs liberal fluids after initial resuscitation show no mortality benefit overall, supporting physiology‑guided, targeted use rather than blanket restriction.

• Role: use as a "triage co‑pilot" to prioritize pharmacist review and stewardship bundles; do not use as an automated hard stop.

• Pharmacist workflow: verify hemodynamics, confirm lack of fluid responsiveness, document rationale, and shift to vasopressor optimization when appropriate.

• High‑risk failure modes needing mandatory human oversight: ongoing resuscitation/hemodynamic instability, active vasopressors, unreliable urine‑output capture, imminent RRT, and baseline‑creatinine uncertainty.

Strengths:

• External validation across two ICU EHR datasets and an interpretable policy tree using bedside variables; signal on MAKE30 in the recommended subgroup.

Limitations:

• Retrospective design with possible unmeasured confounding; marked AUTOC drop on external validation; key operational details and raw subgroup counts missing; no prospective safety or implementation data.

Promising, hypothesis‑generating triage signal for pharmacy‑led IV‑fluid stewardship after AKI in sepsis—pilot in ICU with real‑time KDIGO detection, mandatory clinician checks, and close safety monitoring before scaling.