Low Platelets May Not Predict Severe Outcomes in Newborn Sepsis, Study Finds

A low platelet count at the moment a newborn baby is suspected of developing late-onset sepsis may not, on its own, signal which infants will face the heaviest respiratory burden in the days that follow, according to an exploratory retrospective cohort study published in BMC Pediatrics. The research, led by Xiaochan Yu and colleagues at Xiaolan People’s Hospital of Zhongshan, affiliated with Shantou University Medical College, drew on the neonatal subset of the widely used MIMIC-III critical care database to ask a deceptively simple question: does baseline thrombocytopenia independently predict a predefined composite endpoint of 28-day mortality or persistent mechanical ventilation 48 hours after the onset of presumed infection? The answer the data provided was a cautious and deliberately hedged one, and the way the team arrived at it offers a useful lesson in how confounding can hide inside seemingly straightforward clinical markers.

Late-onset sepsis, typically defined as a serious bloodstream or systemic infection occurring more than 72 hours after birth, remains one of the most feared complications in neonatal intensive care. Unlike early-onset disease, which is usually acquired around the time of delivery, late-onset sepsis emerges in infants who are already hospitalized, often extremely premature and already dependent on life-sustaining technology. Because blood cultures in newborns can be negative even during genuine infection, clinicians frequently rely on a constellation of indirect clues, and thrombocytopenia, a platelet count below 150 × 10⁹ per liter, has long been treated as one of the more ominous among them. Platelets fall during sepsis for several reasons, including consumption in microvascular clotting, impaired production in the bone marrow, and immune-mediated destruction, which is why a dropping count is often read as a signal that infection is severe or worsening.

The problem, as the study’s authors point out, is that platelet counts in the neonatal intensive care unit are rarely a clean signal. Premature infants naturally have lower platelet counts than term babies, and many infants in the NICU are thrombocytopenic for reasons that have nothing to do with sepsis, including neonatal alloimmune thrombocytopenia, drug effects, or the generalized illness of extreme prematurity itself. A low platelet count measured shortly before a suspected sepsis episode may therefore reflect the infant’s baseline fragility rather than the virulence of the infection. If that is the case, then associating thrombocytopenia with poor outcomes tells us less about sepsis severity than about which infants were already the sickest, and any predictive model built on the marker risks simply rediscovering prematurity and baseline respiratory dependence under a different name.

To disentangle this, the researchers constructed a cohort from the neonatal subset of MIMIC-III, a large, de-identified intensive care database developed at Beth Israel Deaconess Medical Center and the Massachusetts Institute of Technology. For each infant, they included the first presumed late-onset sepsis episode that was derived from antibiotic administration linked to blood-specimen collection and occurred more than 72 hours after birth. Baseline thrombocytopenia was defined as a platelet count below 150 × 10⁹ per liter recorded within six hours before the onset of the episode. The endpoint was deliberately designed to capture respiratory burden rather than death alone: an infant met the endpoint if he or she either died within 28 days or remained on mechanical ventilation 48 hours after sepsis onset. This composite reflects the practical reality that in the NICU, prolonged ventilator dependence is one of the most consequential downstream costs of a severe infectious episode.

The cohort ultimately comprised 397 infants, of whom 174, or 43.8 percent, met the composite endpoint. The composition of that endpoint is itself revealing: 172 infants had persistent mechanical ventilation at 48 hours, while only 8 died within 28 days. In other words, the endpoint was overwhelmingly dominated by ventilation status rather than mortality, a fact that would shape every subsequent interpretation. Platelet counts were available for 328 of the infants, and only 28 of those were thrombocytopenic at baseline, a small subgroup that would constrain the statistical precision of every estimate involving the exposure of interest.

The primary analysis used complete-case logistic regression, adjusting for baseline covariates, and produced an adjusted odds ratio of 1.34 for the association between thrombocytopenia and the endpoint, with a 95 percent confidence interval spanning 0.40 to 5.09. That interval is wide enough to include both a substantial protective effect and a substantial harmful one, which in statistical terms means the association was imprecise rather than demonstrably null. The researchers were careful in their language: they did not conclude that platelets are irrelevant, only that this dataset could not yield a precise independent estimate of their contribution. In a field where small, imprecise studies are often overinterpreted in either direction, that restraint is notable.

Far more striking was what happened when the team examined ventilation at the time of sepsis onset. In a contextual sensitivity model that included whether the infant was already being mechanically ventilated when the episode began, baseline ventilation carried an adjusted odds ratio of 30.19, with a 95 percent confidence interval of 13.48 to 73.06, while the platelet estimate barely moved, sitting at an adjusted odds ratio of 1.17 with a confidence interval of 0.27 to 5.49. An odds ratio of that magnitude for baseline ventilation indicates that the composite endpoint was, to a large extent, measuring the infant’s pre-existing respiratory status rather than a new deterioration caused by sepsis. An infant already on a ventilator when infection is suspected is very likely still to be on one 48 hours later, regardless of how the infection evolves, which means the endpoint largely re-encodes baseline illness severity.

Recognizing this, the investigators prespecified a subgroup analysis restricted to infants who were not ventilated at the onset of the episode, a population in which the endpoint would more plausibly reflect new respiratory deterioration attributable to the septic event. This subgroup offered the clearest clinical contrast, but it was also underpowered: 183 infants contributed only 21 endpoint events, yielding an adjusted odds ratio of 1.42 for thrombocytopenia with a confidence interval of 0.17 to 8.32. Once again, the estimate was too imprecise to support any firm conclusion. The authors also conducted a supplementary variable-availability assessment using the paediatric intensive care database from the Children’s Hospital of Zhejiang University School of Medicine, but they explicitly note that this supported only an assessment of variable availability, not a replication of the endpoint analysis, underscoring how difficult it remains to validate such composite outcomes across different data environments.

The methodological honesty of the paper extends to its supplementary materials, which include robustness and contextual ventilation analyses, continuous platelet modeling, diagnostics for maturity variables, a bias-reduced birth-weight sensitivity analysis, and a fully reproducible analytic code archive for cohort construction. The team also validated the recovery of gestational age from the database, an important step given how much of neonatal risk stratification depends on accurate maturity assessment. Model performance was assessed with standard diagnostics, including the Akaike information criterion and the area under the receiver operating characteristic curve, and observed versus predicted endpoint risk was examined by decile. None of these checks rescued a precise platelet signal, but they collectively demonstrate the kind of transparency that exploratory database studies need if their findings are to be interpreted responsibly.

The practical takeaway for clinicians and researchers is twofold. First, in a historical MIMIC-III NICU cohort, baseline thrombocytopenia did not yield a precise independent association with a 48-hour respiratory-support burden endpoint in presumed late-onset sepsis, and the endpoint itself was driven mainly by persistent baseline ventilation, with the thrombocytopenic subgroup too small to support confident inference. Second, and perhaps more importantly, the study illustrates how composite endpoints built around ventilation status can be dominated by pre-existing respiratory dependence, potentially masking or distorting the effects of the exposures they are meant to illuminate. The authors conclude that these findings cannot exclude clinically important platelet associations, and they call for larger, contemporary cohorts using endpoints such as respiratory escalation or ventilator-free days, which are less entangled with baseline ventilator status. Until such studies are available, a low platelet count in a newborn with suspected late-onset sepsis should prompt careful attention, but it should not be treated as a stand-alone verdict on how the next critical 48 hours will unfold.

Subject of Research: Baseline thrombocytopenia as a predictor of respiratory-support burden in presumed neonatal late-onset sepsis

Article Title: Baseline thrombocytopenia and 48-hour respiratory-support burden in presumed neonatal late-onset sepsis: an exploratory MIMIC-III NICU cohort study

Article References: Yu, X., Zheng, Z., Gu, W., & Wei, Z. (2026). Baseline thrombocytopenia and 48-hour respiratory-support burden in presumed neonatal late-onset sepsis: an exploratory MIMIC-III NICU cohort study. BMC Pediatrics. https://doi.org/10.1186/s12887-026-07738-x

Image Credits: AI Generated

DOI: 10.1186/s12887-026-07738-x

Keywords: neonatal sepsis, late-onset sepsis, thrombocytopenia, platelets, NICU, mechanical ventilation, MIMIC-III, retrospective cohort study, neonatology, respiratory support, BMC Pediatrics, clinical prediction