Cost-Effectiveness and Value of Information Analysis of Nutritional Support for Preventing Pressure Ulcers in High-risk Patients: Implement Now, Research Later

Abstract

Background

Pressure ulcers are a major cause of mortality, morbidity, and increased healthcare cost. Nutritional support may reduce the incidence of pressure ulcers in hospitalised patients who are at risk of pressure ulcer and malnutrition.

Objectives

To evaluate the cost-effectiveness of nutritional support in preventing pressure ulcers in high-risk hospitalised patients, and to assess the value of further research to inform the decision to implement this intervention using value of information analysis (VOI).

Methods

The analysis was from the perspective of Queensland Health, Australia using a decision model with evidence derived from a systematic review and meta-analysis. Resources were valued using 2014 prices and the time horizon of the analysis was one year. Monte Carlo simulation was used to estimate net monetary benefits (NB) and to calculate VOI measures.

Results

Compared with standard hospital diet, nutritional support was cost saving at AU$425 per patient, and more effective with an average 0.005 quality-adjusted life years (QALY) gained. At a willingness-to-pay of AU$50,000 per QALY, the incremental NB was AU$675 per patient, with a probability of 87 % that nutritional support is cost-effective. The expected value of perfect information was AU$5 million and the expected value of perfect parameter information was highest for the relative risk of developing a pressure ulcer at AU$2.5 million. For a future trial investigating the relative effectiveness of the interventions, the expected net benefit of research would be maximised at AU$100,000 with 1,200 patients in each arm if nutritional support was perfectly implemented. The opportunity cost of withholding the decision to implement the intervention until the results of the future study are available would be AU$14 million.

Conclusions

Nutritional support is cost-effective in preventing pressure ulcers in high-risk hospitalised patients compared with standard diet. Future research to reduce decision uncertainty is worthwhile; however, given the opportunity losses associated with delaying the implementation, “implement and research” is the approach recommended for this intervention.

Appendix

EVPI [34]

• 1. Assigned probability distributions to the input parameters in the mode as summarised in Table 1 in the main text.

• 2. Sampled random values k times (eg, k = 10,000) from the distributions described above for each intervention.

• 3. Calculated the mean NB for each intervention across all simulations, and identified the preferred baseline decision, that is, the intervention with the maximum expected mean NB ($\max_{i} E_{\theta } {\text{NB}}\left( {i,\theta } \right)$).

• 4. Calculated the NB for each intervention and identified the optimal intervention at each simulation.

• 5. Averaged the NBs from the identified optimal interventions in Step 4 ($E_{\theta } \max_{i} {\text{NB}}\left( {i,\theta } \right)$).

• 6. EVPI per-patient is the difference between the average NBs in steps 5 and 3.

EVPPI [34]

Steps 1–3 as in EVPI algorithm detailed above, then:

• 4. Sampled θ _I once from its joint prior distribution (outer-level simulation).

• 5. Fixed θ _I at their sampled values θ _Ik and simulate the other remaining uncertain parameters θ _Cjk (eg, j = 1,000 times), allowing them to vary according to their conditional probability distribution on θ _I at its sampled value θ _Ik (inner-level simulation).

• 6. Calculated the conditional expected NB of each intervention by evaluating the NB at each (θ _Cjk, θ _Ik) and averaging ($E_{{\left( {\theta_{C} |\theta_{I} } \right)}} {\text{NB}}\left( {i,\theta_{I} ,\theta_{C} } \right)$).

• 7. Identified the intervention that has the highest estimated expected NB given the sampled value for the parameters of interest (θ _Ik) from step 6.

• 8. Repeated steps 4–7 k times (eg, k = 1,000) and calculate the average NB of the preferred interventions identified in step 7 ($E_{{\theta_{I} }} \max_{i} E_{{\left( {\theta_{C} |\theta_{I} } \right)}} {\text{NB}}\left( {i,\theta_{I} ,\theta_{C} } \right)$

• 9. EVPPI is the difference between the average NBs in Steps 8 and 3.

EVSI

To estimate the EVSI for the relative risk (RR) of pressure ulcer with nutritional support (NS) compared to standard care (SC), we assumed that parameters θ _NS and θ _SC represent the probability of pressure ulcers with nutritional support and standard care, respectively. We followed the algorithm adapted from the algorithm reported in Ades 2004 [33].

Steps 0–3 as in EVPI algorithm above, then:

• 4. Simulated the variety of possible results of proposed data collection:

  • 4.1 Drew a sample from the prior distribution of the RR. The logRR ~Normal (μ ₀, τ ₀) where μ ₀ is logRR in the meta-analysis and τ ₀ is its variance.

  • 4.2 Drew a sample baseline parameter θ _SC from its prior distribution: θ _SC~ beta (a, b) where a is the number of patients who developed pressure ulcers and b is the number of patients who did not develop pressure ulcers in the control arm of the meta-analysis.

  • 4.3 Transformed back to obtain an implied prior for θ _NS : θ _NS = θ _SCexp(logRR)

• 5.

  • 5.1 Drew a sample sufficient statistic D, in this case a Binominal numerator, for each arm in the future trialwith size n, assuming equal size arms: r _SC ~ Binomial(θ _SC, n) and r _NS ~ Binomial(θ _NS, n)

  • 5.2 Converted the sufficient statistics to a mean andvariance using the normal approximation:

    μ _D  = log[r _NS × n/r _SC × n],

    τ _D  = [(n − r _SC)/(r _SC × n) + (n − r _NS)/(r _NS × n)]⁻¹

• 6. Updated the prior with the new simulated data to obtain parameters of the posterior distribution:

logRR|D ~ Normal((μ ₀ × τ ₀ + μ _D ,τ _D )/(τ ₀ + τ _D ), τ ₀ + τ _D )

To allow EVSI calculation using TreeAge software, steps 4–6 were performed using Microsoft Excel and repeated 1,000 times generating a ‘coda’ of updated distributions (outer loop). This coda was then transferred to TreeAge for sampling:

• 7. Carried out a nested Monte Carlo simulation (inner loop) drawing 1,000 samples from each posterior distribution of the parameters logRR|D in the Excel coda and from the prior distributions of θ _c, and identified the intervention that has the highest estimated expected NB.

• 8. Calculated the average NB of the preferred interventions identified in Step 7.

• 9. The EVSI is the difference between the average NBs in Steps 8 and 3.