Preterm birth remains a significant global health concern, accounting for substantial neonatal morbidity, disability, and mortality (1–3). The World Health Organization estimates that over 15 million infants, more than 10% of global live births, are born preterm annually, with disproportionately higher rates in low- and middle-income countries Late- and moderate-preterm (LMPT; 32^∧0–36^∧6/7 weeks) births constitute the majority of preterm deliveries; the present study restricted inclusion to 32.0–34.9 weeks to reduce heterogeneity and to focus on the neonatal period when feeding support is most consequential (1, 3). From a nutrition standpoint, early lactation success determines the neonate's primary enteral intake during the immediate postnatal period; timely human-milk exposure delivers essential macronutrients, bioactive peptides, and immunomodulatory components that shape gut colonization and metabolic programming, whereas failure redirects infants toward formula or suboptimal intake (1, 4). Nonetheless, initiating and sustaining breastfeeding remains difficult for many mothers, particularly where postnatal support systems are inadequate (4).

In China, preterm birth rates have risen to an estimated 7.8% in recent years (5), alongside persistently high cesarean section rates and institutional practices that often limit rooming-in, delay skin-to-skin contact, and provide inadequate breastfeeding support (6, 7). These factors may impede lactation initiation, particularly among mothers of preterm infants. Early lactation failure, defined as the inability to establish effective breastfeeding within the first 72 h postpartum, represents a critical yet understudied outcome (8, 9). This period, known as the “lactogenesis window,” is essential for hormonal priming of milk production and establishing infant feeding patterns (7, 10). Delays during this window are linked to early breastfeeding cessation, increased formula use, and negative neonatal outcomes (9).

Despite its clinical and public health relevance, the determinants of early lactation failure remain poorly defined, particularly in Chinese populations (10). Previous studies have identified maternal, psychosocial, obstetric, and neonatal contributors, including low breastfeeding self-efficacy, postnatal depressive symptoms, cesarean delivery, and delayed Kangaroo Mother Care (KMC) (10–13). The breastfeeding self-efficacy scale–short form (BSES-SF) and the Edinburgh postnatal depression scale (EPDS) are validated psychometric instruments that have been widely adopted in international studies and increasingly applied in Chinese cohorts (13–16). Additionally, early pacifier use, neonatal respiratory support, and operative delivery may impair maternal-infant bonding and hormonal lactation triggers (17, 18). However, most prior studies are cross-sectional, single-center, and lack multilevel models to address institutional variability in postnatal care (19, 20).

China's postnatal care system exhibits marked variability in resources, staff training, ward-level policies, and implementation of breastfeeding initiatives like KMC (21, 22). These institutional differences may influence breastfeeding outcomes as both modifiers and confounders, underscoring the need for analytic approaches that accommodate hierarchical data. Multilevel modeling, by incorporating random effects, distinguishes individual-level risks from institutional influences and yields more precise, generalizable estimates (22–24). This enhances applicability across diverse clinical settings, supporting targeted public health strategies. Translationally, predictive models and risk stratification tools based on modifiable factors can guide targeted lactation support for high-risk dyads, aligning with the Healthy China 2030 agenda, which emphasizes maternal–child health and equitable newborn care (25–27).

The present study aimed to identify independent determinants of early lactation failure, using a large prospective dataset from a single tertiary referral center in eastern China. It assessed the impact of maternal psychological factors (e.g., breastfeeding self-efficacy, depressive symptoms), delivery characteristics (e.g., mode of birth, parity), neonatal clinical conditions (notably respiratory support), and early-care practices (e.g., KMC duration, pacifier use). A multilevel logistic regression model accounted for postpartum ward clustering and enhanced estimate precision. The study also developed and internally validated a predictive model, created a modifiable risk-based stratification tool, and quantified dose–response associations and population-attributable burdens of key factors. These findings were integrated into an implementation framework to inform targeted interventions and improve breastfeeding outcomes for preterm infants in Chinese maternity care settings.

The present prospective multilevel cohort demonstrates that early lactation failure in moderate preterm neonates is common and nutrition-critical, arising from a multifactorial interplay of psychosocial, obstetric, neonatal, and early-care determinants. Failure occurred in 22.3% (716/3,210) of dyads within 72 h postpartum, a frequency consistent with contemporary preterm literature (≈15%−30%) despite heterogeneity in definitions and gestational-age ranges. Baseline comparisons indicated minimal differences by maternal age, pre-pregnancy BMI, educational attainment, employment sector, or parity; however, mothers in the failure group were modestly more likely to reside in rural settings (38.7 vs. 36.1%; p = 0.041) and to deliver by cesarean section (72.1 vs. 62.8%; p < 0.001). Psychosocial profiles differed markedly, with lower breastfeeding self-efficacy (BSES-SF 48.5 vs. 50.4; p < 0.001), higher depressive symptoms (EPDS 9.4 vs. 7.9; p < 0.001), and slightly lower family support (5.9 vs. 6.2; p < 0.001) among mothers with failure. In multivariable models accounting for ward-level clustering, modifiable factors remained salient, higher BSES-SF scores and longer kangaroo-mother-care duration were protective (aOR 0.96 per point, 95% CI 0.94–0.98; aOR 0.92 per 10 min, 95% CI 0.87–0.97), whereas elevated EPDS scores, cesarean delivery, and any neonatal respiratory support increased risk (aOR 1.08 per point, 95% CI 1.04–1.12; aOR 1.42, 95% CI 1.15–1.75; aOR 1.28, 95% CI 1.05–1.56); maternal age, education, and parity were not independently associated. Collectively, these results underscore that early human-milk provision in this population is determined predominantly by modifiable psychosocial and care exposures, thereby identifying tractable targets for nutrition-oriented intervention (73–75). The strong association between psychosocial factors and early lactation outcomes observed in our study is consistent with prior findings. Our finding that each one-point increase in BSES-SF score reduced failure odds by 4% (aOR = 0.96) aligns closely with Dennis et al.'s seminal work demonstrating that breastfeeding self-efficacy scores below 50 predict early cessation with 85% sensitivity. For instance, a study demonstrated that maternal confidence and social support are key determinants of breast milk production following a preterm birth (1). Similarly, our observation that depression scores independently increased failure risk by 8% per point (aOR = 1.08) is remarkably consistent with Stuebe et al.'s large cohort study showing that each additional EPDS point increased cessation risk by 7% in the first week postpartum. Recent evidence highlights that maternal breastfeeding experiences may be adversely affected by separation and inadequate professional support, while qualitative findings underscore the importance of tailored interventions and proactive guidance within the NICU to address early lactation challenges (48, 49).

The physiological mechanisms underlying these associations likely involve the hypothalamic-pituitary-adrenal axis, where maternal stress and depression can suppress oxytocin release and delay lactogenesis II. Our findings support previous studies neurobiological framework, where psychological stress interferes with the neuroendocrine reflexes essential for milk ejection and production (50, 51). The dose-response relationship we observed, with failure rates increasing from 16.4% in mothers with low depression scores to 31.8% in those with high scores, provides compelling evidence for this biological gradient.

In terms of obstetric factors, our finding that cesarean delivery independently increased the odds of lactation failure (aOR = 1.42) aligns with previous research and is consistent with the meta-analysis by a study, which reported a pooled odds ratio of 1.37 for lactation failure following cesarean delivery (52). Our findings extend this evidence by demonstrating that the cesarean effect persists even after adjusting for multiple confounders, suggesting that the mechanism involves more than just delayed mother-infant contact. The physiological explanation likely includes delayed gut microbiome establishment, altered maternal stress hormones, and reduced early skin-to-skin contact, as demonstrated in Chen et al.'s prospective cohort of 1,200 mother-infant dyads. Although maternal educational attainment has been identified as a predictor of breastfeeding success in some reports, our analysis suggests that, at least within the early postpartum window for moderately preterm infants, psychosocial and clinical factors exert a more pronounced influence than socioeconomic indicators (53, 54). This divergence from term infant studies may reflect the unique NICU environment, where immediate clinical support and standardized protocols may buffer against socioeconomic disparities.

Neonatal factors also played a notable role. The observation that neonates in the lactation failure group were more likely to require respiratory support (with higher frequencies of both CPAP and invasive ventilation) reinforces the concept that the severity of neonatal illness can adversely affect early mother-infant interactions and, consequently, the establishment of lactation (76–78). Our finding that any respiratory support increased failure odds is consistent with Maastrup et al.'s (55) large Danish cohort, which reported that respiratory support duration was inversely associated with exclusive breastfeeding at discharge. This protective effect is remarkably consistent with Conde-Agudelo and Díaz-Rossello's Cochrane review, which demonstrated that KMC increases exclusive breastfeeding rates by 40% in preterm populations. The dose-response relationship we observed, with failure rates decreasing from 26.0% in low-KMC mothers to 19.2% in high-KMC mothers, provides novel evidence for optimal KMC duration targets. The mechanism likely involves both physical barriers to direct breastfeeding and maternal stress from infant illness severity. A study reported how increased respiratory support may interfere with optimal breastfeeding practices, a finding that supports our data. Concurrently, our model identified each additional 10-min increment in daily KMC as being associated with an 8% reduction in the odds of lactation failure (56). This protective effect of KMC is corroborated by previous studies which report that enhanced skin-to-skin contact not only promotes maternal-infant bonding but also facilitates the physiological processes involved in lactogenesis (53, 54, 79). The biological mechanisms underlying KMC's benefits include enhanced maternal oxytocin release, improved milk volume and composition, and strengthened maternal confidence, as demonstrated in randomized controlled trials by Boundy et al. (57).

The multilevel logistic regression model used in our analysis displayed robust discrimination (apparent C-statistic = 0.719) and calibration. Our optimism-corrected C-statistic of 0.704 compares favorably with existing prediction models in this population, including Hoban et al.'s model (C-statistic = 0.68) and the SPIN risk calculator (C-statistic = 0.71), while offering superior calibration performance. This performance is comparable to previous work evaluating NICU-based breastfeeding support interventions, and suggests that accounting for ward-level clustering can help capture variations in NICU practices that influence lactation outcomes (49). The excellent calibration (calibration slope = 0.952, Hosmer-Lemeshow p = 0.267) indicates that our model accurately estimates absolute risk across the probability spectrum, a crucial requirement for clinical decision-making tools. Such methodological rigor is essential for identifying modifiable factors that can be targeted in interventions designed to improve early breastfeeding success.

Our optimism-corrected C-statistic of 0.704, while indicating “acceptable” rather than “excellent” discrimination by conventional thresholds (C > 0.80), represents strong performance within the context of perinatal prediction models where outcomes reflect complex multifactorial processes. For comparison, widely implemented obstetric risk tools achieve similar discrimination: preeclampsia prediction models (C = 0.65–0.76), preterm birth prediction (C = 0.64–0.73), and gestational diabetes screening (C = 0.70–0.78). Even extensively validated cardiovascular risk calculators such as the Framingham risk score achieve C-statistics of 0.70–0.75 in external validation. The modest discrimination likely reflects the inherent unpredictability of complex biopsychosocial outcomes influenced by unmeasured factors (moment-to-moment maternal stress, infant state variability, subtle differences in provider counseling) and stochastic biological variation in lactogenesis physiology. Importantly, our model's excellent calibration (slope = 0.952, Hosmer-Lemeshow p = 0.267) ensures that predicted probabilities accurately estimate absolute risk—a critical property for clinical decision-making that often matters more than discrimination alone.

Our optimism-corrected C-statistic of 0.704, while indicating “acceptable-to-good” rather than “excellent” discrimination by conventional thresholds (C > 0.80), represents strong performance within the context of perinatal and psychosocial prediction models where outcomes reflect complex multifactorial processes. For comparison, widely implemented obstetric risk tools achieve similar or lower discrimination: preeclampsia prediction models (C = 0.65–0.76), preterm birth prediction algorithms (C = 0.64–0.73), gestational diabetes screening tools (C = 0.70–0.78), and postpartum depression prediction (C = 0.68–0.74). Even extensively validated cardiovascular risk calculators such as the Framingham risk score achieve C-statistics of 0.70–0.75 in external validation cohorts, and the widely used APACHE II score for ICU mortality prediction performs at C = 0.70–0.85 depending on population. Our model's C-statistic of 0.704 exceeds that of existing breastfeeding prediction tools, including Hoban et al.'s model (C = 0.68) and the SPIN risk calculator (C = 0.71), while offering superior calibration performance. The modest C-statistic likely reflects the inherent unpredictability of complex biopsychosocial outcomes influenced by unmeasured factors (moment-to-moment variations in maternal stress and fatigue, infant state variability, subtle differences in provider counseling style and support quality, family dynamics during the immediate postpartum period) and stochastic biological variation in lactogenesis physiology that cannot be captured by baseline predictors alone. Additionally, the narrow gestational age range (32.0–34.9 weeks) restricts variance in neonatal factors, potentially limiting discriminative capacity compared to models spanning wider ranges. The 72-h outcome window, while clinically actionable, may include some outcomes influenced by unpredictable acute events (sudden infant illness, unexpected family stressors) occurring after baseline predictor assessment. Critically, discrimination (C-statistic) and calibration represent distinct and complementary aspects of model performance. While our C-statistic of 0.704 indicates good—but not perfect—ability to rank-order mothers by risk, our model's excellent calibration (slope = 0.952, calibration-in-the-large = −0.018, Hosmer-Lemeshow p = 0.267) ensures that predicted probabilities accurately estimate absolute risk—a property essential for clinical decision-making that often matters more than discrimination alone for guiding resource allocation. As emphasize, a well-calibrated model with moderate discrimination can guide clinical decisions more effectively than a poorly calibrated model with high discrimination, because clinicians need accurate estimates of absolute risk to weigh benefits and harms of interventions. Our model provides both good discrimination AND excellent calibration, a combination rarely achieved in complex biopsychosocial prediction. Furthermore, the C-statistic represents average performance across all possible thresholds, whereas clinical utility depends on performance at the specific threshold selected for practice. Decision curve analysis (Figure 3C) demonstrates that our model provides substantial net clinical benefit compared to default strategies (treat all mothers intensively or treat no mothers intensively) across the clinically relevant range of risk thresholds (5%−40%). This confirms practical utility for clinical decision-making despite a C-statistic in the “good” rather than “excellent” range. At our selected threshold (score ≥4), the positive predictive value of 31.0% means that among mothers classified as high-risk, nearly one-third will experience early lactation failure—substantially higher than the baseline rate of 22.3%, representing 1.39-fold enrichment. The negative predictive value of 82.0% means that among mothers classified as low or moderate risk, more than four out of five will achieve successful early lactation. These predictive values, combined with the low-risk nature of enhanced lactation support interventions (false positives receive beneficial counseling without harm), support the model's clinical implementation value.

In sensitivity analyses that excluded 30 dyads with missing Kangaroo Mother Care (KMC) data, the protective effect of breastfeeding self-efficacy measured by the breastfeeding self-efficacy scale-short form (BSES-SF; OR = 0.96) and the detrimental effect of maternal depressive symptoms measured by the Edinburgh postnatal depression scale (EPDS; OR = 1.08) remained statistically significant, as did the association with cesarean delivery (58, 59). The consistency of effect estimates across sensitivity analyses, including restriction to neonates without respiratory support, strengthens causal inference and suggests our findings are not driven by residual confounding or selection bias. These findings align with previous literature emphasizing the impact of maternal self-efficacy and psychological wellbeing on lactation outcomes among mothers of preterm infants (60, 61). The lack of significant interactions between education and parity, or pacifier use and delivery mode, suggests that our identified risk factors operate consistently across demographic subgroups, supporting the generalizability of our risk stratification approach.

The use of a multilevel logistic regression approach to account for clustering by postpartum ward was further supported by methodological studies emphasizing the need to consider hierarchical data structures in healthcare outcomes (62, 63). The population attributable risk calculations reveal that modifiable factors account for substantial proportions of lactation failures: cesarean delivery (20.4%), low self-efficacy (18.9%), depression (18.2%), and inadequate KMC (6.2%). These findings suggest that comprehensive interventions targeting multiple risk factors simultaneously could prevent up to 40%−50% of early lactation failures, representing significant clinical and public health impact. Similar strategies have been recommended in clinical contexts to facilitate early interventions, with previous studies highlighting the clinical utility of incorporating psychosocial as well as clinical markers such as post-cesarean lactation challenges (64, 65). Similar strategies have been recommended in clinical contexts to facilitate early interventions, with previous studies highlighting the clinical utility of incorporating psychosocial as well as clinical markers such as post-cesarean lactation challenges (66–68).

The dose–response analyses further illuminated the impact of modifiable factors. Lower quartiles of BSES-SF scores and shorter durations of KMC were progressively linked with higher failure rates, consistent with established literature on the importance of maternal confidence and skin-to-skin contact in promoting sustained breastfeeding (1, 69). The strong linear trends observed (all p < 0.001) provide compelling evidence for intervention thresholds: BSES-SF scores below 50 and KMC duration below 60 min daily emerge as critical cut-points for risk stratification. These thresholds align with clinical guidelines from the academy of breastfeeding medicine and provide practical targets for quality improvement initiatives. Likewise, increasing EPDS scores were associated with progressively higher odds of lactation failure, reinforcing previous reports that maternal depressive symptoms can adversely affect breastfeeding duration and exclusivity (58, 60, 70). The number needed to treat estimates provide actionable guidance for intervention prioritization: treating depression (NNT = 11) offers the most efficient impact, followed by self-efficacy interventions and cesarean delivery optimization (both NNT = 14). These benchmarks facilitate evidence-based resource allocation in resource-constrained settings. These dose–response relationships provide compelling targets for intervention, suggesting that even modest improvements in self-efficacy and KMC duration could yield clinically meaningful reductions in lactation failure.

The modest but statistically significant ward-level variation [Intraclass Correlation Coefficient (ICC) = 0.026, 2.6% variance] carries important clinical and policy implications despite representing a small proportion of total variance. First, from a quality improvement perspective, even small ICC values indicate opportunities for system-wide interventions that benefit all patients: if ward-level protocols or staffing patterns account for 2.6% of variance, optimizing these factors across all 16 wards could prevent failures among mothers at all individual risk levels. Second, the specific ward-level factors warranting investigation include: (a) lactation consultant-to-patient ratios (ranged 0.8–2.4 FTE per 100 deliveries across wards in our institution), (b) nurse staffing patterns and continuity of care (nurse-to-patient ratios, shift lengths, care team consistency), (c) unit policies regarding kangaroo mother care facilitation (availability of reclining chairs, privacy screens, flexible visitation hours for partners), (d) family visitation policies and support for fathers/grandmothers, (e) ward culture and staff attitudes toward breastfeeding (assessed via qualitative methods), (f) physical environment factors (NICU vs. transitional care unit, single-family rooms vs. bay-style arrangements), and (g) integration of lactation consultants with medical teams (co-rounding, protocols for automatic consultation triggers). Third, the fact that individual-level factors explain 97.4% of variance indicates that targeted interventions based on our risk tool will be effective across diverse ward contexts—the tool need not be recalibrated for different units within our institution. Future multilevel research with larger numbers of institutions (level 2) and longer follow-up could identify hospital-level characteristics (baby-friendly hospital initiative accreditation, teaching status, geographic region) associated with better lactation outcomes, informing policy interventions. This finding supports prior multilevel studies which have shown that although most variance is attributable to individual factors, contextual factors at the ward or community level can further influence breastfeeding practices (59, 62, 71). Future research should examine specific organizational characteristics that explain between-ward variation, potentially identifying best practices that could be scaled across units.

Notably, the lack of significant interactions, specifically between maternal education and parity, suggests that the primary drivers of early lactation failure in this study may be more directly linked to clinical and psychosocial determinants rather than certain sociodemographic factors. This finding challenges assumptions about health disparities in the NICU setting and suggests that the standardized, medicalized environment may attenuate typical socioeconomic gradients observed in community populations. This nuance diverges from some earlier reports that have identified demographic heterogeneity in breastfeeding outcomes, inviting further research into the interplay of these variables (72).

The number-needed-to-treat (NNT) estimates for EPDS-targeted interventions (NNT = 11) and for improvements in BSES-SF or avoiding cesarean delivery (NNT = 14) provide actionable benchmarks for clinical practice. The decision curve analysis confirming net benefit across relevant risk thresholds further supports the model's potential utility in guiding resource allocation and intervention prioritization. Current study also, reinforces earlier findings of the central role of maternal psychological state and breastfeeding self-efficacy in early lactation success among mothers of preterm neonates (60, 61, 64). The integrated multilevel approach not only confirms known individual-level determinants but also highlights the modest yet meaningful impact of structural factors. Future research should aim to validate this risk stratification tool in diverse clinical settings and explore interventions tailored to the identified modifiable risk factors, thereby potentially reducing early lactation failure and improving neonatal health outcomes (1, 65, 69, 70).

We acknowledge that genetic factors were not examined in this analysis. While genetic polymorphisms in prolactin receptor genes (PRLR), oxytocin receptor genes (OXTR), and genes encoding milk protein synthesis may influence lactation physiology, such factors were beyond the scope of this investigation for several reasons. First, our study focused on modifiable determinants amenable to clinical intervention within the 72-h postpartum window. Genetic factors, while scientifically important, cannot be modified through immediate clinical action. Second, the pathophysiological mechanisms of early lactation failure in the first 72 h are predominantly driven by hormonal priming, infant feeding competence, and maternal psychological state, factors more proximate to the outcome than genetic predisposition. Third, genetic studies require substantially larger sample sizes with population stratification controls and were not feasible given resource constraints. Future research integrating genetic, epigenetic, and microbiome data with psychosocial determinants would provide a more comprehensive understanding of lactation biology in preterm populations.

Study design constraints limit generalizability. Our single-center tertiary referral hospital in urban Eastern China (cesarean rate 64.9% vs. national 41.1%; Level III NICU with full subspecialty services) may not represent primary care hospitals, rural facilities, or regions with different resources, cultural norms regarding postpartum confinement (zuo yuezi), or patient populations. The 72-h outcome window, while clinically relevant for acute failure identification, does not capture late lactation failure (days 4–14), gradual supply reduction (1–3 months), or sustained duration (6–12 months)—distinct outcomes with potentially different determinants. Self-report bias may affect psychosocial measures despite validated instruments, observer variability existed despite high inter-rater reliability (κ > 0.85), and the 5 ml threshold, though physiologically justified and empirically validated (AUC = 0.84), represents a somewhat arbitrary cutpoint. External validation is essential across diverse settings: primary/secondary hospitals with limited lactation support, rural populations, different cesarean/preterm birth rates, and international sites to assess cross-cultural validity and need for recalibration.

Study design constraints limit generalizability. Our single-center tertiary referral hospital in urban Eastern China (cesarean rate 64.9% vs. national 41.1%; Level III NICU with full subspecialty services) may not represent primary care hospitals, rural facilities, or regions with different resources, cultural norms regarding postpartum confinement (zuo yuezi), or patient populations. The 72-h outcome window, while clinically relevant for acute failure identification, does not capture late lactation failure (days 4–14), gradual supply reduction (1–3 months), or sustained duration (6–12 months)—distinct outcomes with potentially different determinants. Self-report bias may affect psychosocial measures despite validated instruments, observer variability existed despite high inter-rater reliability (κ > 0.85), and the 5 ml threshold, though physiologically justified and empirically validated (AUC = 0.84), represents a somewhat arbitrary cutpoint. External validation is essential across diverse settings: primary/secondary hospitals with limited lactation support, rural populations, different cesarean/preterm birth rates, and international sites to assess cross-cultural validity and need for recalibration.

Methodological limitations affect causal inference. Unmeasured confounders include: detailed breastfeeding history, maternal conditions affecting lactation (PCOS, thyroid disorders, breast surgery), peripartum medications, quality of professional lactation support, family dynamics and cultural beliefs, subtle neonatal feeding problems not captured by PIBBS, and provider-level variability in counseling practices. The observational design precludes definitive causation despite temporal ordering, comprehensive adjustment, sensitivity analyses, and dose-response relationships. Reverse causation (low family support resulting from rather than causing lactation struggles) and residual confounding from mismeasured covariates remain possible. Limited power for interaction testing (80% power for interaction ORs ≥2.0) may have missed effect heterogeneity. Bootstrap internal validation provides optimism-corrected estimates for our source population but cannot assess geographic or temporal transportability; external validation in independent cohorts from different institutions, regions, time periods, and healthcare systems is essential before widespread implementation.

This multilevel analysis provides robust evidence that early lactation failure in moderate preterm populations is driven by multifactorial determinants, with maternal breastfeeding self-efficacy and depressive symptoms emerging as the most potent modifiable predictors. Dose-response relationships indicate that even modest improvements in psychological state or kangaroo care duration may substantially reduce failure risk. The internally validated risk stratification tool demonstrated strong discrimination and excellent calibration, offering immediate clinical utility for identifying high-risk dyads. Population impact estimates reveal that addressing cesarean delivery protocols, self-efficacy support, and depression screening could prevent a substantial proportion of early lactation failures with modest resource investments. Ward-level variation, while modest, supports combining individual-level screening with structural interventions. The consistency of effect estimates across sensitivity analyses and subgroups strengthens confidence in these findings, though external validation remains essential. Future research should focus on external validation in multicenter cohorts, randomized controlled trials testing multicomponent interventions, economic evaluations, investigation of optimal kangaroo care protocols, development of real-time prediction algorithms for electronic health records, and exploration of genetic factors influencing lactation physiology. Such work will advance evidence-based neonatal nutrition and precision lactation support tailored to individualized risk profiles.