29.2 Critical Care Research Methods, Statistics, IRB Oversight & Quality Improvement

Reading an ICU Trial

Effect measures:

• Absolute risk reduction (ARR) $=$ control event rate $-$ treatment event rate.
• Relative risk reduction (RRR) $= \text{ARR} / \text{control event rate}$.
• Number needed to treat (NNT) $= 1 / \text{ARR}$.
• Odds ratio approximates relative risk only when the outcome is uncommon.
• Hazard ratio describes the instantaneous rate over time and comes from time-to-event analysis.

Worked example: mortality falls from 40% to 32%. The ARR is 8 percentage points, the RRR is 20%, and the NNT is $1/0.08 = 12.5$, so about 13 patients must be treated to prevent one death. The same 20% relative reduction applied to a 2% baseline mortality gives an ARR of 0.4 percentage points and an NNT of 250. Always convert to absolute terms.

Design vocabulary:

• Randomization balances known and unknown confounders; allocation concealment prevents gaming assignment; blinding prevents differential care and outcome assessment.
• Intention-to-treat analyzes patients as randomized and preserves the benefit of randomization; it is conservative for superiority trials. Per-protocol analysis is biased by differential dropout but is needed to support noninferiority claims, where a sloppy trial can create a false appearance of equivalence.
• Noninferiority margin: the largest acceptable difference, specified in advance. A result is noninferior only if the confidence interval excludes that margin.
• Composite outcomes can be driven by their least important component; look at each element.
• Surrogate endpoints (oxygenation, a biomarker) often fail to predict patient-centered outcomes, as inhaled nitric oxide in ARDS demonstrated.
• Subgroup analyses are hypothesis-generating unless pre-specified with an interaction test.
• Stopping early for benefit overestimates effect size.
• Pragmatic and cluster-randomized designs, stepped-wedge trials, and platform trials with Bayesian adaptive randomization (REMAP-CAP) are increasingly common; understand that adaptive designs report posterior probabilities rather than single p values.
• External validity: screened-to-enrolled ratios, exclusion criteria, and usual care in the control arm determine whether the result applies to your patient.

Statistical concepts:

• Type I error (alpha) is a false positive; type II error (beta) is a false negative. Power $= 1 - \beta$.
• Confidence intervals convey precision and plausible effect sizes; a wide interval crossing 1.0 cannot exclude meaningful benefit or harm.
• A p value above 0.05 does not demonstrate equivalence; many ICU trials are underpowered for mortality.
• Choose tests appropriately: t test for two normally distributed means, ANOVA for more than two, Mann-Whitney U or Wilcoxon for non-normal data, chi-square or Fisher exact for proportions, log-rank for survival curves, and logistic or Cox regression for adjustment.
• Mortality definitions matter: ICU, hospital, 28-day, 60-day, and 90-day mortality are not interchangeable, and discharge practices distort hospital mortality.

Diagnostic Test Characteristics

This is why a D-dimer is useful in low-probability patients (high sensitivity, high negative predictive value in low prevalence) and useless in high-probability patients, and why screening a low-prevalence ICU population generates false positives.

Human Subjects Protection

• IRB review is required for research involving human subjects. The IRB evaluates risk-benefit balance, subject selection equity, informed consent adequacy, privacy and confidentiality protections, and monitoring plans.
• Core documents and principles: the Nuremberg Code, the Declaration of Helsinki, and the Belmont Report (respect for persons, beneficence, justice), implemented in the United States through the Common Rule (45 CFR 46) and FDA regulations.
• Consent in the ICU: most patients lack capacity, so legally authorized representative consent is used, with deferred consent from the patient on recovery where permitted.
• Exception from informed consent (EFIC) for emergency research requires all of:
  1. A life-threatening condition for which available treatments are unproven or unsatisfactory.
  2. Prospective consent is not feasible within the therapeutic window.
  3. The research offers the prospect of direct benefit.
  4. Community consultation and public disclosure before and after the trial.
  5. An independent data monitoring committee.
  6. Attempts to contact a representative within a defined therapeutic window.
• Vulnerable populations (pregnant people, prisoners, children, those with impaired decision-making) have additional protections.
• Conflicts of interest must be disclosed; clinical trial registration before enrollment and reporting of results are expected.
• Quality improvement versus research: QI aimed at improving local care with existing knowledge generally does not require IRB review, but systematic investigation designed to produce generalizable knowledge does. When in doubt, ask the IRB.

Quality Improvement and Patient Safety

Methods:

• Model for improvement: aim, measures, changes, then Plan-Do-Study-Act cycles.
• Lean reduces waste; Six Sigma reduces variation; high-reliability organization principles emphasize preoccupation with failure and deference to expertise.
• Measures: structure (staffing, equipment), process (bundle compliance, time to antibiotics), outcome (mortality, CLABSI rate), and balancing measures that detect unintended harm.
• Run charts and statistical process control (control) charts distinguish common cause variation from special cause variation; a sustained shift of points on one side of the median signals real change. A simple before-after comparison with a p value cannot separate a trend from an intervention effect.
• Risk-adjusted benchmarking (APACHE, SAPS, SOFA-based models) allows comparison across units, with observed-to-expected mortality ratios.

Safety tools:

• Checklists and bundles (central line insertion, ventilator, sepsis) reduce harm when coupled with culture change, as the Keystone ICU project demonstrated.
• Root cause analysis is retrospective; failure mode and effects analysis is prospective.
• Incident reporting and a just culture that distinguishes human error, at-risk behavior, and reckless behavior.
• Structured handover (for example, I-PASS) reduces communication-related error, the leading contributor to sentinel events.
• Rapid response systems and early warning scores target deterioration outside the ICU.
• Error taxonomy: a near miss reaches no patient, an adverse event causes harm, and a sentinel event causes severe harm or death and triggers mandatory review. A never event (such as wrong-site surgery or a retained foreign body) should never occur.

ICU Organization and Administration

• Staffing and structure: high-intensity physician staffing models and daytime intensivist coverage are associated with better outcomes; mandatory overnight in-house intensivist staffing did not improve mortality in a randomized trial. Nurse-to-patient ratios and the presence of clinical pharmacists and respiratory therapists affect outcomes.
• Tele-ICU extends specialist oversight to units without on-site intensivists.
• Triage and capacity: prioritize by likely benefit, use objective admission and discharge criteria, and plan for surge with crisis standards of care developed in advance, including a triage team separate from bedside clinicians.
• Reimbursement and documentation: critical care time billing requires documented time devoted to the critically ill patient; severity of illness documentation affects risk adjustment and benchmarking.
• Regulatory context: CMS Conditions of Participation, Joint Commission standards, public reporting, and value-based purchasing programs that penalize certain hospital-acquired conditions and readmissions.
• Burnout and the clinician workforce: high prevalence in critical care, with moral distress a specific driver; mitigation includes workload management, ethics support, and team debriefing.