Probiotics for infection prevention in preterm neonates

Abbreviations

Abbreviation

Definition

Note

BEmONC

Basic emergency obstetric and neonatal care

Operationalized as facilities without C-section capabilities

CEmONC

Comprehensive emergency obstetric and neonatal care

Operationalized as facilities with capabilities to perform C-section

PAF

Population Attributable Fraction

Intervention Overview

Research Background

Preterm infants are susceptible to a serious disease of the gastrointestinal tract called necrotizing enterocolitis (NEC). While the pathogenesis of NEC is not entirely understood, it is hypothesized that the paucity of normal enteric bacteria and the delayed onset of bacterial colonization in preterm infants relative to term infants leave preterm infants susceptible to NEC infection. Probiotic supplements have been shown to reduce the risk of NEC incidence among preterm infants in a 2014 Cochrance review [Cochrane-Review-Probiotics] and a recent meta-analysis from 2025 [Lee-Him-et-al-2025].

The Cochrane review found that “probiotics supplementation significantly reduced the incidence of severe NEC (stage II or more) (typical relative risk (RR) 0.43, 95% confidence interval (CI) 0.33 to 0.56; 20 studies, 5529 infants) and mortality (typical RR 0.65, 95% CI 0.52 to 0.81; 17 studies, 5112 infants). There was no evidence of significant reduction of nosocomial sepsis (typical RR 0.91, 95% CI 0.80 to 1.03; 19 studies, 5338 infants).” [Cochrane-Review-Probiotics]

The 2025 review found: “Probiotic supplementation reduced the risk of NEC by 61% (95% CI: 51–69%), the risk of all-cause neonatal mortality by 25%(95% CI: 8–39%), and the risk of invasive infection by 19% (95% CI: 9–28%) in preterm newborns, when compared to control. There were significant subgroup differences in the risk of NEC by probiotic type. The risk of NEC was reduced by all types of probiotics; however, the greatest reduction was found for Bifidobacteriumspp. by 84% (95% CI: 69–92%), and the smallest reduction was found for Saccharomyces spp. by 7% (95% CI: 55% reduction to 94% increase), though this was not statistically significant when compared to control or placebo. However, there were no significant differences in mortality or invasive infection by probiotic type.”” [Lee-Him-et-al-2025]

The Gates Foundation has specifically been interested in probiotic supplementation with Bifidobacterium infantis (B. infantis), which has been found to be a particularly effective probiotic strain for reducing NEC incidence [Lee-Him-et-al-2025]. However, given there were no significant differences in effect on invasive infection by probiotic type, we will use the summary effect of all probiotics on invasive infection as our modeled effect of probiotic supplementation on sepsis mortality among preterm infants for our simulation.

Vivarium Modeling Strategy

This section describes how a probiotic-treatment intervention can be implemented and calibrated for the MNCNH Portfolio model.

Affected Outcomes

Outcome

Effect

Modeled?

Note (ex: is this relationship direct or mediated?)

Neonatal sepsis and other neonatal infections Mortality Probability \(\text{CSMRisk}_i^\text{sepsis}\)

Adjust multiplicatively using RR

Yes

Baseline Coverage Data

Baseline coverage for probiotics availability at health facilities should be 0% for both BEMONC and CEMONC facilities.

Baseline Coverage of Neonatal Probiotics (placeholder values)

Birth Facility

Coverage Mean (%)

Coverage Distribution (%)

Notes

Home Birth

0

N/A

BEmONC Facilities

0

N/A

CEmONC Facilities

0

N/A

Vivarium Modeling Strategy

This intervention requires adding an attribute to all simulants to specify if a neonate has access to a facility with access to probiotics. Since the neonatal mortality model does not explicitly represent incidence of sepsis, we will not track explicitly if a simulant receives probiotics. Instead the model will have different cause-specific mortality rates for sepsis for individuals with and without access to probiotics (implemented with a slightly confusing application of our Risk and RiskEffect components from vivarium_public_health).

In order to be eligible for this intervention, simulants must be in the early neonatal age group and born preterm (<37 weeks old). This eligibility is based on an impact table for Bifidobacterium infantis (B. infantis) provided to us by the BMGF team.

digraph probiotics { rankdir = LR; nodesep = 1.0; ranksep = 1.2; node [shape=box] rankdir = LR; birth [label="Simulant is born"] preterm [label="Simulant is preterm (<37 weeks)"] not_preterm [label="Simulant is not preterm"] preterm_wo [label="Simulant is eligible but does not have access to probiotics"] preterm_w [label="Simulant is eligible and has access to probiotics"] not_preterm_wo [label="Simulant is not eligible for probiotics"] birth -> preterm birth -> not_preterm preterm -> preterm_wo preterm -> preterm_w not_preterm -> not_preterm_wo }

The Risk component adds an attribute to each simulant indicating whether the simulant (if in the target population) has access to probiotics during the neonatal period, which we assume to be 0.0% regardless of birth facility choice in our baseline scenario. births in BEmONC facilities have lower access than CEmONC facilities.

To make this work naturally with the RiskEffect component, it is best to think of the risk as “lack of access to probiotics”. With this framing, the RiskEffect component requires data on (1) the relative risk of sepsis mortality for people with lack of access to probiotics, and (2) the population attributable fraction (PAF) of sepsis due to lack of access to probiotics. We will use the decision tree below to find the probability of sepsis mortality with and without access to probiotics that are logically consistent with the baseline delivery facility rates and baseline probiotics coverage.

In Vivarium, this risk effect will modify the sepsis mortality pipeline, resulting in

\[\text{CSMRisk}_i^\text{sepsis} = \text{CSMRisk}^\text{sepsis}_{\text{BW}_i, \text{GA}_i} \cdot (1 - \text{PAF}_\text{no probiotics}) \cdot \text{RR}_i^\text{no probiotics}\]

where \(\text{RR}_i^\text{no probiotics}\) is simulant i’s individual relative risk for “no probiotics”, meaning \(\text{RR}_i^\text{no probiotics} = \text{RR}_\text{no probiotics}\) if simulant i accesses a facility without probiotics, and \(\text{RR}_i^\text{no probiotics} = 1\) if simulant i accesses a facility with probiotics.

If there are other interventions also affecting the CSMR of sepsis, the pipeline will combine these effects, and we can write out the math for this risk explicitly as

\[\text{CSMRisk}^\text{sepsis}_{i, \text{updated}} = \text{CSMRisk}^\text{sepsis}_{i, \text{original}} \cdot (1 - \text{PAF}_\text{no probiotics}) \cdot \text{RR}_i^\text{no probiotics}\]

This reduces to the previous formula if there are no other interventions, and we would have

\[\text{CSMRisk}^\text{sepsis}_{i, \text{original}} = \text{CSMRisk}^\text{sepsis}_{\text{BW}_i, \text{GA}_i}\]
Risk Effect Parameters for Lack-of-Access-to-probiotics

Parameter

Mean

Source

Notes

\(\text{RR}^\text{no probiotics}\)

\(1/\text{RR}^\text{probiotics}\)

N/A

Value to be used in sim

\(1/\text{RR}^\text{probiotics}\)

RR = 0.81 (95% CI: 0.72 to 0.91). Parameter uncertainty implemented as a lognormal distribution: get_lognorm_from_quantiles(0.81, 0.72, 0.91)

[Lee-Him-et-al-2025]

PAF

see below

see below

see Calibration strategy section below for details on how to calculate PAF that is consistent with RR, risk exposure, and facility choice model

Calibration Strategy

The following decision tree shows all of the paths from delivery facility choice to probiotics availability. Distinct paths in the tree correspond to disjoint events, which we can sum over to find the population probability of sepsis mortality. The goal here is to use internally consistent conditional probabilities of sepsis mortality for the subpopulations with and without access to probiotics, so that the baseline scenario can track who has access to probiotics and still match the baseline sepsis mortality rate.

digraph probiotics { rankdir = LR; facility [label="Facility type"] home [label="p_sepsis_without_probiotics"] BEmONC [label="probiotics?"] CEmONC [label="probiotics?"] BEmONC_wo [label="p_sepsis_without_probiotics"] BEmONC_w [label="p_sepsis_with_probiotics"] CEmONC_wo [label="p_sepsis_without_probiotics"] CEmONC_w [label="p_sepsis_with_probiotics"] facility -> home [label = "home birth"] facility -> BEmONC [label = "BEmONC"] facility -> CEmONC [label = "CEmONC"] BEmONC -> BEmONC_w [label = "available"] BEmONC -> BEmONC_wo [label = "unavailable"] CEmONC -> CEmONC_w [label = "available"] CEmONC -> CEmONC_wo [label = "unavailable"] }

\[\begin{split}\begin{aligned} p(\text{sepsis}) &= \sum_{\text{paths without probiotics}} p(\text{path})\cdot p(\text{sepsis}|\text{no probiotics})\\ &+ \sum_{\text{paths with probiotics}} p(\text{path})\cdot p(\text{sepsis}|\text{probiotics})\\[.1in] p(\text{sepsis}|\text{no probiotics}) &= \text{RR}_\text{no probiotics} \cdot p(\text{sepsis}|\text{probiotics}) \end{aligned}\end{split}\]

where \(p(\text{sepsis})\) is the probability of dying from sepsis in the general population, and \(p(\text{sepsis}|\text{probiotics})\) and \(p(\text{sepsis}|\text{no probiotics})\) are the probability of dying from sepsis in setting with and without access to probiotics. For each path through the decision tree, \(p(\text{path})\) is the probability of that path; for example the path that includes the edges labeled BEmONC and unavailable occurs with probability that the birth is in a BEmONC facility times the probability that the facility has probiotics available.

When we fill in the location-specific values for delivery facility rates, probiotics coverage, relative risk of mortality with probiotics access, and mortality probability (which is also age-specific), this becomes a system of two linear equations with two unknowns (\(p(\text{sepsis}|\text{probiotics})\) and \(p(\text{sepsis}|\text{no probiotics})\)), which we can solve analytically.

As mentioned above, it is convenient to model this intervention like a dichotomous risk factor, so that we can reuse the Risk and RiskEffect components in Vivarium Public Health, rather than having to write new components from scratch. Calling the intervention a risk factor can sound a bit confusing because intervention access is a good thing, so it doesn’t sound “risky.” Instead, we flip it so the risk factor is “lack of access to the intervention.” The RiskEffect component expects a relative risk (RR) and a population-attributable fraction (PAF). Because we are flipping the direction of the risk factor, we need to use the inverse of our original RR, so:

\[\text{RR}_{\text{no intervention}} = \frac{1}{\text{RR}_{\text{intervention}}}\]

The PAF is the proportion of deaths due to the outcome that would not occur if all births had access to the intervention. Since we use the equation \(p(\text{outcome}|\text{intervention}) = (1 - \text{PAF}_\text{no intervention}) \cdot p(\text{outcome})\) in the RiskEffect component, we solve for \(\text{PAF}_\text{no intervention}\) as follows:

\[\text{PAF}_{\text{no intervention}} = 1 - \frac{p(\text{outcome}|\text{intervention})}{p(\text{outcome})}\]

where the terms on the right hand side can be obtained by solving the system of equations above.

Here is some pseudocode for deriving the PAF and RR of “lack of access to the intervention”:

p_sepsis = neonatal_sepsis_mortality_risk
relative_risk = 1/RR_probiotics # this represents the RR of lack of access to probiotics

p_sepsis_probiotic = p_sepsis / (
(p_home * (1 - p_probiotic_home) * relative_risk)
+ (p_home * p_probiotic_home)
+ (p_BEmONC * (1 - p_probiotic_BEmONC) * relative_risk)
+ (p_CEmONC * (1 - p_probiotic_CEmONC) * relative_risk)
+ (p_BEmONC * p_probiotic_BEmONC)
+ (p_CEmONC * p_probiotic_CEmONC)
)
paf_no_probiotic = 1 - (p_sepsis_probiotic / p_sepsis)

The above strategy was used for the implementation of this intervention model in the MNCNH portfolio. Note that the value of “p_sepsis” is arbitrary and does not directly affect the resulting PAF value. Documentation for the implemented strategy above and the simplified versiow below (including links to relevant parameters used) are both shown for reference.

\[ \begin{align}\begin{aligned}p_\text{intervention} = \sum_\text{facility type} p_\text{facility type} * p_{\text{intervention} | \text{facility type}}\\E(\text{RR}_\text{no intervention}) = p_\text{intervention} + (1 - p_\text{intervention}) * \text{RR}_\text{no intervention}\\\text{PAF}_\text{no intervention} = \frac{E(\text{RR}) - 1}{E(\text{RR})}\end{aligned}\end{align} \]

Where,

PAF calculation parameters

Parameter

Definition

Value

Note

\(p_\text{facility type}\)

Proportion of population that delivers in a given facility type

Defined in the Overall delivery setting rate section of the Facility choice model document

\(p_{\text{intervention} | \text{facility type}}\)

Proportion of eligible population in a giving facility type that receives the intervention at baseline

Defined in the Baseline Coverage Data section of this document

Scenarios

Todo

Describe our general approach to scenarios, for example set coverage to different levels in different types of health facilities; then the specific values for specific scenarios will be specified in the MNCNH Portfolio model.

Assumptions and Limitations

  • We assume that probiotics availability captures actual use, and not simply the treatment being in the facility

  • We assume that the delivery facility is also the facility where preterm infants will receive prophylactic probiotic supplementation

  • We assume that the relative risk of sepsis mortality with probiotics in practice is a value that we can find in the literature

  • We do not specifically measure the impact of this intervention on NEC, the condition directly affected by the intervention, because it is not modeled by GBD. We use the less specific neonatal sepsis and other neonatal infections cause in GBD instead.

  • We assume that the effect of probiotics on invasive infection is an appropriate proxy for the effect on sepsis cause-specific mortality.

  • The effect size of probiotics on all-cause mortality is greater in magnitude than the effect on invasive infection from [Lee-Him-et-al-2025]. This is inconsistent with the causal hypothesis we are modeling in our simulation. However, there is less statistical precision in the effect estimate for all-cause mortality than for invasive infection, so we accept this inconsistency.

Validation and Verification Criteria

  • Population-level mortality rate should be the same as when this intervention is not included in the model

  • The ratio of sepsis deaths per birth among those without probiotics access divided by those with probiotics access should equal the relative risk parameter used in the model

  • The baseline coverage of probiotics in each facility type should match the values in the artifact

  • Check whether the effect size on all cause mortality is within the confidence interval of the observed effect from [Lee-Him-et-al-2025]

References