RENOVATE Trial

Or… an introduction to Bayesian learning

Ed Palmer

ST8 in Anaesthesia & Intensive Care Medicine

2025-05-23

Learning Objectives / Housekeeping

  • Develop an intuition for Bayesian statistics.
  • Appreciate the differences between Bayesian and frequentist reasoning.
  • Unlearn a few bad statistical habits!
  • No maths. No formulas*. Accessible to all.
  • Interactive. Feel free to interrupt.
  • No singling out, but volunteers welcome.
  • Some statistical concepts are so ingrained, keep an open mind.

Let’s Challenge Our Assumptions

  • An RCT investigated the effect of steroids in septic shock on mortality.
  • Mortality was 43% in the treatment group and 49.1% in the control group (p = 0.03).
  • What is the correct interpretation of this p value?

Let’s Challenge Our Assumptions

  • An RCT investigated the effect of steroids in septic shock on mortality.
  • The relative risk of mortality in the treatment group was 0.88 (95% confidence interval, 0.78 to 0.99).
  • What is the correct interpretation of this 95% confidence interval?

My Prediction

  • No more will get the answers than would be expected by random chance.
  • These are hard questions, but foundational.
  • These concepts come from “Frequentist statistics”.
  • Bayesian statistics conceptualises probabilities differently.
  • Most frequentist statistics are interpreted incorrectly; already Bayesian?
  • If you are confused. That’s ok.

Setting

  • 33 Brazilian hospitals
  • Nov 2019 - Nov 2023
  • HFNO vs NIV for ARF

Background

  • Number of different approaches to managing ARF
  • Pharmacological:
    • Bronchodilators
    • Antibiotics
    • ..
    • Oxygen
  • Non-pharmacological
    • Positioning
    • Sputum clearance
    • Mechanical support: IMV, NIV/CPAP, HFNO

Background

Non-invasive Ventilation

  • Delivers a pressure differential
  • Can titrate O2 occording to need
  • Augments minute volume
  • Applies PEEP
  • Limits interventions, often poorly tolerated
  • Need breaks to eat etc.

High Flow Nasal Oxygen

  • Provides high inspiratory flow in excess of peak inspiratory flow
  • Bulk flow reduces deadspace
  • Comfortable; patients wear it for longer.
  • Heated and humidified
  • ?PEEP; skeptical

Background

  • Both are useful tools when used correctly.
  • Dynamic evidence landscape
  • Infective aetiologies and NIV?
  • RECOVERY-RS: CPAP
  • FLORALI: HFNO
  • HERNANDEZ: HFNO

Population

  • Inclusion Criteria:
    • Aged \(\geq\) 18 y/o admitted to hospital (ED/ICU/Wards) with ARF
    • Spo2 <90% or PaO2 < 7.9 kPa RA
    • Signs of increased WOB or tachypnoea >25 bpm

Population

  • Exclusion Criteria:
    • Urgent need for endotracheal intubation
    • Prolonged respiratory pauses
    • Cardiorespiratory arrest
    • GCS \(\leq\) 12
    • Heart rate < 50 bpm with decreased level of consciousness
    • Arterial blood pH < 7.15
    • Haemodynamic instability
    • Contraindication to NIV (vomiting, secretions, GCS < 12, pneumothorax)
    • Do not intubate order
    • NIV use in ACPO prior to randomisation

Population

  1. Nonimmunocompromised ARF
  2. Immunocompromised ARF
  3. Acute Cardiogenic Pulmonary Oedema (ACPO)
  4. COPD
  5. COVID-19

Intervention

  • HFNO
    • Airvo-2, Fisher & Paykel Healthcare
    • Flow:
      • 30L/min in COPD, 45L/min otherwise
      • Titrated up to 60L/min as tolerated
    • FiO2:
      • Started at 0.5
      • COPD: titrated for SpO2 88-92%
      • Other groups: 92-98%

NIV rescue therapy was allowed for COPD and ACPO only.

Control

  • NIV (face mask)
    • Pressure:
      • COPD: IPAP 12-16 cmH2O, EPAP 4 cmH2O
      • Other groups: IPAP 12-14 cmH2O, EPAP 8cmH2O
      • Maximum IPAP 20 cmH2O EPAP of 12cmH2O
    • Tidal volume:
      • 6-9 ml/Kg* of ideal body weight
    • FiO2: as for intervention

24hr use of NIV encouraged.

Outcome

  • Primary: Endotracheal intubation (or death) within 7 days of randomisation
    • Criteria for intubation standardised
  • Secondary:
    • 28 + 90 day mortality
    • Mechanical ventilation-free days at 28 days
    • ICU-free days at 28 days
  • Tertiary:
    • ICU + Hospital LOS within 90 days
    • Vasopressor-free days within 28 days
    • The proportion of patients who received a do-not-intubate order within 7 days after randomisation

Key Methods

  • Non-inferiority unblinded bayesian adaptive randomised controlled trial.

  • 5 distinct groups.

  • “Adaptive”; dynamic borrowing.

  • 1:1 randomisation, permuted block size, allocation concealment.

  • Interim analyses: futility, non-inferiority, or superiority.

  • Composite outcome.

Key Methods

  • Non-inferiority unblinded bayesian adaptive randomised controlled trial.

  • 5 distinct groups.

  • “Adaptive”; dynamic borrowing.

  • 1:1 randomisation, permuted block size, allocation concealment.

  • Interim analyses: futility, non-inferiority, or superiority.

  • Composite outcome.

Non-inferiority Trial

  • Most RCTs are “superiority” trials; is one treatment “better” than the other?
  • Is this always necessary?
  • Is one treatment less toxic or less expensive?
  • One can be “better” if clinical efficacy the same.
  • How do you prove that two treatments are the same?
  • Can anyone recognise the issue?

It’s About precision

  • Without enough data, all comparisons are non-inferior

Let’s go Bayesian

  • A simulated trial
  • Population: 10 Adults from the general population admitted to the ICU with a diagnosis of septic shock.
  • Intervention: “Vaso-pushin”™; 5 patients
  • Control: Standard care; 5 patients
  • Outcome: Increase in MAP. MCID increase of 10 mmHg assuming baseline MAP of 55.

Borrowing

  • Let’s conduct a thought experiment
  • Consider an RCT in ARF conducted in 2024 and another in 2025.
  • How might information transfer from one to the other?
  • Could we include the controls of the 2024 study as “controls” in the 2025 study?
  • Why might we want to do this?

Borrowing

  • Bayesian approach naturally allows for borrowing
  • Caution must be exercised
  • Are the groups really similar? Are they “exchangeable”?
  • Increased precision
  • But… results will gravitate towards the group mean

Trial Results

Stopping Criteria

  • April 2021 (1st interim); immunocompromised. Futility. n = 51
  • March 2023 (5th interum); COVID-19, non-inferiority. n = 895
  • Oct 2023 (6th interum); Non-immunocomprimise, non-inferiority. n = 501
  • Oct 2023 (6th interum); Cardiogenic pulmonary oedema, non-inferiority. n = 274
  • Final analyis: COPD. n = 79

Conclusions

Conclusions

  • Met pre-specified criteria for non-inferiority for 4 of 5 cohorts
  • But… results not robust to strong modelling assumptions (borrowing)
  • Need further study: COPD, immunocompromised and ACPO

Stengths

  • Large, multi-centred RCT.
  • Bayesian; transparent.
  • Auhtors conclusions are well made.
  • Broad generalisability for the patients I see.
  • Others…

Weaknessess

  • Late (but necessary) inclusion of COVID-19 to trial protocol.
  • Modification of trial protocol to include death as composite outcome.
  • Very small numbers for COPD ?selection bias.
  • Over/inappropriate borrowing from dissimilar groups; failed sensitivity analysis.
  • Criteria for futility too harsh.
  • High cross over from HFNO to NIV in COPD.
  • Others…

My Take Aways

  • You can’t bayesian your way out of a bad idea.
  • Reasurring and has equipoise for ARF.
  • Has challenged my assumptions for ACPO.
  • I would still use NIV first line in COPD, and CPAP in ACPO.
  • And what about yours…

Other resources