Phenotypic characteristics associated with preterm births in the Indian population

Abstract

Background

Preterm birth (PTB), defined as the birth of an infant before 37 weeks of gestation, poses significant health risks, including increased mortality and long-term disabilities. India stands as a major contributor to global PTB-related mortality, primarily due to its substantial prevalence. The complex temporal relationship among multiple etiological factors makes predicting PTB occurrence challenging. This is further complicated by the diverse PTB phenotypic categories, namely maternal, fetal, placental, parturition, and pathway-to-delivery features, which affect birth outcomes.

Methods

Our study focused on PTB in the Indian population using a comprehensive dataset from the GARBH-Ini cohort. We analyzed 21 variables across maternal, fetal, placental, parturition-related, and other phenotypic categories in the context of Phenotype I (PTB/Non-PTB) and Phenotype II, which classifies PTB based on the pathway to delivery as either Spontaneous Onset of Labor (SOL) or Caregiver-Initiated (CGI). We assessed the prevalence of these phenotypic variables and their co-occurrences using the complete dataset. We applied hierarchical clustering to identify key PTB-associated phenotypic patterns and their impact on neonatal outcomes, including NICU admissions and early neonatal mortality. The optimal number of clusters (K = 20) was selected based on silhouette scores. Clusters with intermediate PTB composition (40–80% PTB, “grey zone”) were excluded to minimize misclassification and improve interpretability. Additionally, we used one-year follow-up data to examine infant-related phenotypic variables and their distribution across PTB and non-PTB cases.

Results

We identified key phenotypic variables across maternal, fetal, placental, and parturition categories that significantly influenced PTB outcomes. In SOL cases, fetal (perinatal sepsis, polyhydramnios) and parturition-related (PROM, short cervix, peripartum bleeding) features were significantly associated. In contrast, CGI cases were predominantly associated with maternal complications like illness suggestive of birth asphyxia, hypertension, and hypothyroidism. Placental abnormalities, such as placenta previa, in combination with PROM and short cervix, were notably prevalent in PTB cases, impacting neonatal mortality, NICU admissions, and infant health.

Conclusions

Compared with non-PTB births, PTB cases showed higher prevalence of several key phenotypic features, including short cervix (10.84% vs 3.42%), perinatal sepsis (16.58% vs 3.63%), previous PTB among multiparous women (19.21% vs 7.1%), amniotic fluid abnormalities, PROM (21.67% vs 13.3%), and hypertensive disorders including eclampsia (9.52% vs 5.17%) (all P_adj < 0.05). Stratified analyses demonstrated that SOL-associated PTB was enriched for short cervix, perinatal sepsis, PROM, and amniotic fluid abnormalities, whereas CGI-associated PTB was primarily linked to pre-eclampsia. PTB was also associated with higher NICU admissions and early neonatal mortality. This detailed PTB phenotyping enables healthcare providers to better monitor and manage high-risk pregnancies, leading to improved maternal and neonatal outcomes.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12884-026-09836-1.

Introduction

Currently, 15 million infants are born premature, of which a million unfortunate succumb to complications of prematurity [1]. Many survivors face the risk of a lifetime of disability, including learning disabilities and visual and hearing problems. Based on the countries with reliable trend data, PTB rates have not declined despite socioeconomic development [2]. It is estimated that over 60% of PTB worldwide comes from Sub-Saharan Africa and South Asia [3], and India alone is the most significant contributor to the global burden [4, 5]. While the definition of PTB is on a time scale, defined as a live birth < 37 weeks of gestation, the characteristics of this adverse event are heterogeneous. The diversity of etiologic factors and events that occur during the perinatal period, collectively referred to as a ‘phenotype,’ contributes to the heterogeneity of PTB and influences neonatal and infantile outcomes [6, 7].

In this context, a framework has been proposed to classify PTB phenotypes based on maternal, fetal, and placental conditions, signs of parturition, and the PTB pathway to delivery [8]. A cross-sectional study of the INTERGROWTH-21st utilized this framework to classify and identify PTB phenotypes [9]. They identified 12 PTB phenotypes associated with different patterns of neonatal outcomes. Such types and distribution of PTB phenotypes might vary from population to population.

In this study, we have identified PTB phenotypes using maternal, fetal, placental conditions, and pathway to delivery, as these groups may have distinct underlying causes, risk factors, and clinical trajectories [10]. Given that pregnancy is influenced by a multitude of factors, leading to significant variations between pregnancies, identifying meaningful subgroups requires approaches that can capture this multidimensional heterogeneity. Conventional supervised approaches, such as logistic regression, are primarily designed for outcome prediction and rely on predefined response variables, which may not fully capture latent phenotypic subgroups in a heterogeneous syndrome like PTB. Unsupervised clustering provides a data-driven framework for uncovering naturally occurring phenotypic patterns and co-occurring clinical features without imposing prior, outcome-driven assumptions. Therefore, we adopted a hierarchical clustering approach, which enabled us to compare smaller groups of similar pregnancies and identify key phenotypic categories (Table S1) and their combinations critical to PTB outcomes. By uncovering these distinct clinical patterns, phenotypic characterization provides the granular data necessary to refine predictive models and develop targeted, personalized prenatal interventions tailored to the specific risk profile of each pregnancy subgroup.

Methods

Study design and data collection

For the present study, secondary data available in the GARBH-Ini cohort were used. This hospital-based pregnancy cohort commenced in May 2015 at Gurugram Civil Hospital, Gurugram, Haryana, India. Women were enrolled before 20 weeks of pregnancy and followed till delivery and once post-partum. For each participant, a wide array of data was collected, including maternal anthropometric measurements, clinical and obstetric histories, socioeconomic information, various biospecimens, and maternal and fetal biometry ultrasound assessments. The detailed methodologies of the GARBH-Ini study have been documented in a prior publication [11].

The institutional ethics committees of Gurugram Civil Hospital (Gurugram), Safdarjung Hospital (New Delhi), the Translational Health Science and Technology Institute, and the Indian Institute of Technology Madras approved the study.

Dataset preparation and phenotypic variable definition

We analyzed 21 variables (maternal, fetal, placental features, signs of parturition, and others) from the GARBH-Ini cohort (Table S1). These variables were selected for their relevance to the phenotypic classification framework proposed by Villar et al. [8] and for their critical role in analyzing their relationships with PTB outcomes. The data were obtained from 10,494 participants enrolled in the cohort between 2015 and 2021. For this study, only participants who had delivered at the Gurugram Civil Hospital and had available details of delivery and pregnancy outcomes (N = 5,231) were selected. Stillbirths and those who had delivered at < 28 weeks gestational age (numbers were too small to make inferences with certainty, Fig. 1A) were excluded.

Fig. 1.

Overview of data selection and clustering workflow used for PTB phenotyping. A A stepwise data selection process was used to construct the final dataset (N = 3640) for clustering analysis. This included the application of inclusion/exclusion criteria to phenotypic variables from the GARBH-Ini cohort to obtain a complete-case subset for clustering. B Clustering workflow applied to the selected dataset, including preprocessing, distance metric selection, hierarchical clustering, and cluster labeling. This workflow was used to identify clinically meaningful subgroups based on phenotypic patterns associated with PTB

Exploratory data analysis

We compared the proportion of PTB and non-PTB cases for each variable using a one-tailed Fisher’s exact test, with the alternative hypothesis set to ‘greater.’ Since several categorical comparisons involved low numbers of observations in one or more categories, one-tailed Fisher’s exact test was used instead of the chi-square test. P-values were adjusted for multiple comparisons using the Bonferroni method. A variable was considered significantly more prevalent in PTB cases if the P_adj value was < 0.05. In addition, we assessed pairwise associations between variables by calculating the proportion of total cases in the PTB and non-PTB groups where one variable was recorded as ‘yes,’ given that another variable was also recorded as ‘yes.’

To delineate the etiological heterogeneity of PTB, we stratified our analysis by the pathway to delivery: Spontaneous Onset of Labor (SOL), in which labor began naturally without medical intervention, and Caregiver-Initiated (CGI), in which delivery was initiated by clinical decision due to maternal or fetal indications. This stratification enables the identification of phenotype patterns and risk factors that may be unique to each pathway, thereby providing a more nuanced understanding of PTB in our cohort.

Hierarchical clustering

Based on the exploratory assessment (Figures S1, S2), variables with less than 20% missing data and a prevalence of at least 5% were selected for the clustering analysis. Only complete case records were used for clustering analyses. Participants with missing values in the selected variables were excluded, and no imputation strategies were applied. Variables with lower prevalence were excluded, as they typically contribute little to meaningful cluster differentiation. Based on these criteria, eight variables were selected (five maternal and one each from the fetal, placental, and other categories; Table S2). Additionally, two continuous numerical variables, cervix length and amniotic fluid index, were included, as their continuous nature provides finer granularity and can enhance the clustering process by capturing subtle differences between data points. Rows with missing values in the selected variables were removed, resulting in a complete-case dataset comprising 3640 participants. This included 370 (10.16%) PTB cases and 3,270 (89.83%) non-PTB cases and was used for PTB phenotype analysis. It also included 3,096 SOL cases (85.05%) and 111 CGI cases (3.04%), which were subsequently used to analyze PTB phenotypes based on delivery pathway. PTB was recorded in 292 (9.43%) of the SOL cases and 21 (18.91%) of the CGI cases.

Agglomerative (bottom-up) hierarchical clustering, using both binary and continuous variables, was performed to identify distinct clusters of PTB and non-PTB cases. Given the imbalance between PTB and non-PTB cases, we conducted 100 random samplings, selecting an equal number of cases from each group in every iteration to enhance the robustness of our analysis (Fig. 1B). Since our dataset contained both binary and continuous numeric variables, we used the Gower distance metric [12], which normalizes variables based on their respective ranges, thereby ensuring comparability across data types. This metric functions similarly to the Manhattan distance, which is defined as the sum of the absolute differences between the coordinates of two points. The Gower metric functions similarly, but incorporates range normalization, making it suitable for mixed data.

We used silhouette scores [13] ranging from −1 to + 1 to evaluate clustering quality, with higher values indicating more distinct, well-defined clusters (Supplementary Information; Figure S5a). By analyzing the distribution of average silhouette scores across varying numbers of clusters (K) and their sizes, we identified K = 20 as the optimal solution, yielding an average silhouette score of ~ 0.6 across all three datasets used for Phenotype I (also referred to as PTB Phenotype) and Phenotype II (corresponding to the SOL and CGI subset) analysis. The average median cluster sizes were 13.82 for the full dataset (N = 3640), 10.56 for the SOL subset (N = 3096), and 1.32 for the CGI subset (N = 111). Clustering was implemented in R using the distmix() function from the kmed package [14] (v0.4.2) and the hclust() function from the stats package (v4.1.2).

Classification of the clusters

A cluster with at least 5 cases was labeled a PTB cluster if 80% or more of the cases were PTB cases. Similarly, a cluster with at least 5 cases was labelled a non-PTB cluster if less than 40% of its cases were PTB (Figure S5b). Clusters that did not meet either of these conditions were considered “grey zone” (Fig. 1) and excluded from further analysis, as their mixed composition could introduce ambiguity and undermine the interpretability of our comparisons by increasing the risk of misclassification. On average, in 100 random samples, the majority of clusters fell into this grey zone: about 14 out of 20 clusters in the full dataset, 16 out of 20 in the SOL dataset, and 18 out of 20 in the CGI dataset. This filtering step allowed us to retain optimally sized clusters for statistical testing and compare the prevalence of variables between PTB and non-PTB clusters.

Assessing the prevalence of phenotypic variables in PTB clusters

After identifying PTB and non-PTB clusters and their corresponding enrolment IDs, all 21 variables were mapped (Table S1) across all clusters in 100 random samples. We first assessed the normality of the distribution of the proportion of cases marked ‘yes’ for each feature. Because the distributions deviated from normality and contained extreme values, the assumptions for parametric t-tests were not met. Since the distributions were non-normal, we applied a one-sided Wilcoxon test to determine whether the proportion of cases marked ‘yes’ was higher in PTB than in non-PTB cases. The null hypothesis (H₀) assumed that there was no difference between the groups, meaning the median proportions were equal. To account for multiple comparisons, the P_adj values were calculated using the Bonferroni method, implemented in R through a custom in-house script. Effect size was quantified using Fold Change (FC), calculated as the ratio of the proportion in the PTB group (foreground) to that in the non-PTB group (background). Features were prioritized based on FC and P_adj < 0.05 threshold, ranking them according to their relevance to PTB outcomes. A feature was considered overrepresented in PTB clusters if its P_adj < 0.05. Additionally, we extended this analysis to assess pairwise feature combinations.

Analyzing the newborn features with the variables of PTB

The relationship between preterm infant features and the selected variables, independent of the clustering-based analysis, was studied. Through pairwise comparisons, infant features were examined in relation to maternal, fetal, placental, and parturition characteristics. For each pair, we counted the number of PTB cases in which both the infant feature and the phenotypic variable were marked as ‘yes.’ These cases formed the foreground group, representing the co-occurrence of the two features. The background group comprised the remaining PTB cases where either one or neither of the features was marked as ‘yes.’ A one-sided Fisher’s exact test (alternative = “greater”) was used to evaluate whether co-occurrence was significantly enriched compared to the background. FC was calculated as the ratio of the proportion of cases in the foreground to that in the background. Bonferroni correction was applied to account for multiple testing within each dataset. This analysis was performed separately for the full dataset (N = 5,129), the SOL subset (N = 4,324), and the CGI subset (N = 172), enabling the identification of both consistent and subset-specific associations between infant outcomes and phenotypic features of PTB.

Results

The baseline characteristics of participants included and excluded from this study are provided in Table 1 and Table S3.

Table 1.

Baseline characteristics of the participants included in the study

Sociodemographic characteristics	Selected Participants (N = 5,231) set median (IQR) or N (%) or mean ± SD
Age (years)	24 (21, 26)
BMI at enrolment into the cohorta
Underweight	24.90%
Normal	58.47%
Overweight	12.59%
Obese	2.42%
Hemoglobin (g/dL)	9.2 (8.6–10.4)
Weight (kgs)	58
Socioeconomic statusb
Upper class	0.26%
Upper middle class	11.01%
Lower middle class	26.34%
Upper lower class	38.48%
Lower class	0.82%
Undetermined	23.07%
Level of education
Illiterate	16.97%
Literate or primary school	8.96%
Middle school	16.34%
High school	19.02%
Post-high school diploma	21.46%
Graduate	14.01%
Post-graduate	3.21%
Unknown	0.00%
Occupation
Unemployed	91.35%
Unskilled worker	3.80%
Semi-skilled worker	1.62%
Skilled worker	2.46%
Clerk, shop, farm owner	0.30%
Semi-professional	0.22%
Professional	0.21%
Unknown	0.00%
History of any chronic illnessesc
Absent	97.03%
Present	2.98%
Parity
Nulliparous	39.11%
Multiparous	60.86%

^aPre-pregnancy BMI (kg/m²) was calculated from participants’ weight (kg) and height (m) measured at enrolment. BMI categories were defined as underweight (< 18.5), normal (18.5–24.9), overweight (25.0–29.9), and obese (≥ 30.0)

^bSocioeconomic status was assessed using Modified Kuppuswamy’s socioeconomic scale, calculated using the education and occupation of the head of the family and monthly family income

^cChronic illnesses include a history of hypertension, diabetes, cardiac disease, and thyroid disorders

Of the 5,231 selected participants, 609 (11.63%) experienced PTB. Pathway-to-delivery information was available for 4,656 participants (88.99%). Among them, 4390 (94.24%) had SOL, while the remaining 175 had CGI deliveries. Among those with SOL, 472 (10.75%) reported PTB, whereas 31 (1.77%) in the CGI group did. Exploratory data analysis was conducted on all the selected variables (Table S1) to assess their distributions and the proportion of missing data (Figures S1, S2, S4).

We also identified features related to newborn conditions using follow-up information available for 5,129 participants (98.03%). Of these, 579 (11.28%) were PTB. The infant condition-related features are presented in Table S4, and the results of the exploratory analysis are shown in Figure S3.

Prevalence of characteristics in the PTB phenotype

Pre-eclampsia, short cervix, hypertension, perinatal sepsis, history of previous PTB among multiparous women, polyhydramnios, oligohydramnios, and PROM were significantly more prevalent in the PTB group than in the non-PTB group. The prevalence of all characteristics in the PTB phenotype is summarized in Table 2.

Table 2.

Enrichment of phenotypic variables in PTB versus non-PTB cases in the complete dataset (PTB-Phenotype, N = 5,231). The table presents the number and percentage of PTB and non-PTB cases for each phenotypic variable, grouped by clinical category. Statistical significance indicates overrepresentation of the variable in PTB cases compared to non-PTB cases, based on P_adj values corrected (Bonferroni method) for multiple testing. ** indicates P_adj < 0.01; * indicates P_adj < 0.05; variables without asterisks are not statistically significant

Class	Phenotypic Variables	# PTB (%)	# Non-PTB (%)	Padj value
Maternal	Short cervix **	66 (10.84%)	158 (3.42%)	< 0.0001
Maternal	Chorioamnionitis	6 (0.99%)	37 (0.8%)	1.00
Maternal	Bacterial vaginosis	35 (5.75%)	300 (6.49%)	1.00
Maternal	Hepatitis	9 (1.48%)	41 (0.89%)	1.00
Maternal	Respiratory infections	50 (8.21%)	377 (8.15%)	1.00
Maternal	UTI	197 (32.35%)	1,522 (32.92%)	1.00
Maternal	Eclampsia **	58 (9.52%)	239 (5.17%)	< 0.0001
Maternal	Medical Disorders	423 (69.46%)	3,096 (66.97%)	1.00
Maternal	Anemia	277 (45.48%)	1,932 (41.79%)	0.685
Maternal	Hyperthyroidism *	13 (2.13%)	35 (0.76%)	0.0382
Maternal	Hypothyroidism	31 (5.09%)	148 (3.2%)	0.212
Maternal	Parity	380 (62.4%)	2804 (60.65%)	1.00
Maternal	Perinatal sepsis **	101 (16.58%)	168 (3.63%)	< 0.0001
Maternal	PTB History **	73 (19.21%)	199 (7.1%)	< 0.0001
Maternal	Uterine Rupture	2 (0.33%)	6 (0.13%)	1.00
Fetal	FGR	163 (26.77%)	1819 (39.35%)	1.00
Fetal	Fetal Anomaly	4 (0.66%)	10 (0.22%)	0.281
Fetal	Oligohydramnios **	10 (1.64%)	3 (0.06%)	< 0.0001
Fetal	Polyhydramnios *	13 (2.13%)	35 (0.76%)	0.0102
Placental	Placenta Previa	356 (58.46%)	3057 (66.13%)	1.00
Parturition	PROM **	132 (21.67%)	615 (13.3%)	< 0.0001
Other	Biomass Fuel	65 (10.67%)	395 (8.54%)	0.0503

A stratified analysis by SOL and CGI deliveries was conducted to assess the association between the delivery pathway and PTB. Results highlighted contrasting phenotypic profiles across the two subtypes (Table 3). In the SOL subset, the prevalence of perinatal sepsis, short cervix, previous history of PTB, polyhydramnios, oligohydramnios, and PROM was significantly higher in the PTB group. In contrast, in the CGI subset, only pre-eclampsia was significantly higher in the PTB group.

Table 3.

Enrichment of phenotypic variables in PTB cases stratified by pathway to delivery. A total of 4390 participants had SOL, and 175 had CGI deliveries. The table shows the number and percentage of PTB and non-PTB cases for each phenotypic variable within the SOL and CGI subgroups. Statistical significance reflects the overrepresentation of the variable in PTB cases relative to non-PTB cases within each subgroup, as determined by P-values adjusted for multiple testing (Bonferroni method). ** indicates P_adj < 0.01; * indicates P_adj < 0.05; variables without asterisks are not statistically significant

Class	Phenotypic Variables	SOL				CGI
		# PTB (%)	# Non-PTB (%)	Padj value	# PTB (%)		# Non-PTB (%)	Padj value
Maternal	Short Cervix	61 (12.92%) **	141 (3.6%)	< 0.0001	1 (3.23%)		5 (3.47%)	1.00
Maternal	Chorioamnionitis	5 (1.06%)	36 (0.92%)	1.00	1 (3.23%)		0 (0%)	1.00
Maternal	Bacterial vaginosis	23 (4.87%)	267 (6.81%)	1.00	4 (12.9%)		5 (3.47%)	0.804
Maternal	Hepatitis	9 (1.91%)	36 (0.92%)	0.701	0 (0%)		3 (2.08%)	1.00
Maternal	Respiratory infections	38 (8.05%)	324 (8.27%)	1.00	3 (9.68%)		11 (7.64%)	1.00
Maternal	UTI	159 (33.69%)	1,301 (33.21%)	1.00	11 (35.48%)		49 (34.03%)	1.00
Maternal	Pre-eclampsia	27 (5.72%)	174 (4.44%)	1.00	13 (41.94%) *		21 (14.58%)	0.0174
Maternal	Anemia	227 (48.09%)	1658 (42.32%)	0.145	6 (19.35%)		47 (32.64%)	1.00
Maternal	Hypertension	5 (1.06%)	23 (0.59%)	1.00	2 (6.45%)		5 (3.47%)	1.00
Maternal	Hypothyroidism	14 (2.97%)	111 (2.83%)	1.00	3 (9.68%)		3 (2.08%)	1.00
Maternal	Parity	291 (61.65%)	2,288 (58.4%)	1.00	12 (38.71%)		77 (53.47%)	1.00
Maternal	Perinatal Sepsis	80 (16.95%) **	143 (3.65%)	< 0.0001	5 (16.13%)		10 (6.94%)	1.00
Maternal	History of previous PTB	53 (18.21%) **	153 (6.69%)	< 0.0001	5 (41.67%)		6 (7.79%)	0.0851
Maternal	Uterine Rupture	1 (0.21%)	3 (0.08%)	1.00	0 (0%)		0 (0%)	1.00
Fetal	FGR	122 (25.85%)	1,577 (40.25%)	1.00	13 (41.94%)		64 (44.44%)	1.00
Fetal	Fetal Anomaly	4 (0.85%)	6 (0.15%)	0.0653	0 (0%)		0 (0%)	1.00
Fetal	Oligohydramnios	8 (1.69%) **	2 (0.05%)	< 0.0001	1 (3.23%)		1 (0.69%)	1.00
Fetal	Polyhydramnios	11 (2.33%) **	25 (0.64%)	0.0039	0 (0%)		2 (1.39%)	1.00
Placental	Placenta Previa	287 (60.81%)	2,632 (67.18%)	0.997	15 (48.39%)		81 (56.25%)	0.841
Parturition	PROM	107 (22.67%) **	529 (13.5%)	< 0.0001	11 (35.48%)		38 (26.39%)	0.209
Other	Biomass Fuel	50 (10.59%)	333 (8.5%)	0.0779	4 (12.9%)		15 (10.42%)	0.444

Cluster-specific enrichment of key phenotypic variables in the subtypes of PTB

Hierarchical clustering-based enrichment analysis identified distinct phenotypic determinants of PTB (Figure S6, Table S5) and for phenotypes stratified by delivery pathway (Figs. 2A, 2B, Table S5).

Fig. 2.

Distribution of the percentage of ‘Yes’ responses for phenotypic variables in PTB and non-PTB clusters using the SOL and CGI subsets. The figure shows the distribution of ‘Yes’ responses for the 21 phenotypic variables across clusters derived from two subsets of the dataset used in SOL and CGI analysis: (A) Spontaneous onset of labor (SOL; N = 3096) and (B) Clinician-initiated delivery (CGI; N = 111). (C) Distribution of significantly overrepresented co-occurring phenotypic variable pairs in PTB clusters within the SOL subset, selected for co-occurrence analysis based on variables identified as significant in panels A and B. Green denotes PTB clusters, and purple denotes non-PTB clusters. The distributions highlight differences in the prevalence of clinical phenotypic variables between PTB and non-PTB groups within each subset. P_adj values for multiple comparisons are shown alongside each distribution to indicate the significance of observed differences

In the PTB phenotype, the clusters exhibited significant enrichment of maternal, fetal, and parturition-related complications. Perinatal sepsis emerged as the most enriched variable, observed in 82.18% of PTB clusters versus 12.68% in non-PTB clusters (p < 0.001). Pre-eclampsia was associated with a pronounced increase in PTB clusters (17.47% versus 1.67%, p < 1e-17), as well as other maternal features, such as hypothyroidism (3.63% versus 1.92%, p < 0.001). Rare complications like uterine rupture (2.92%) and oligohydramnios (4.58% versus 0.10% p < 2.08e-27) were almost exclusively observed in PTB clusters, underscoring their clinical significance despite low prevalence. We also observed that PROM, short cervix, and placental abnormalities, such as placenta previa, were less common in PTB cases.

In the SOL-PTB phenotype, perinatal sepsis remained the most prominent feature, enriched in 84.39% of PTB clusters compared to 16.44% in non-PTB clusters (p < 1e-22). Cervical insufficiency emerged as a key driver, with short cervix significantly overrepresented in PTB clusters (12.25% versus 7.56%, p < 0.001), alongside PROM (40.76% versus 17.77%, p < 0.001). Amniotic fluid abnormalities, such as oligohydramnios, showed a substantial enrichment (7.53% versus 0.11%, p < 0.001), while maternal infections, such as UTIs, were more frequent in PTB cases (76.39% versus 50.94%, p < 0.001).

In the CGI-PTB phenotype, the enrichment profile revealed hypertension (15.56% versus 0%, p < 0.001), FGR (100% versus 2.38%, p < 0.001), hypothyroidism (18.25% versus 6.27%, p < 0.001), and respiratory infections (13.49% versus 0%, p < 0.001) as the prominent factors for PTB.

Pairwise associations

An analysis of pairwise associations between phenotypic variables in PTB and non-PTB cases (Table S6) and in the SOL and CGI subsets (Tables S7, S8) was conducted to identify combinations of clinical features that co-occur more frequently in PTB. For the PTB phenotype, several pairs of characteristics showed more than twofold higher prevalence compared to non-PTB cases. For example, anemia co-occurred with a short cervix in 10.11% of PTB versus 3.31% of non-PTB cases; with perinatal sepsis in 14.8% versus 3.47%, and with a PTB history in 11.55% versus 4.4%. Previous history of PTB occurred more frequently with perinatal sepsis (21.92% versus 2.01%), and short cervix (8.22% versus 4.02%) in PTB than in non-PTB groups. Other important pairs seen in the PTB group include hypertension with PROM (30.77% versus 11.43%); placenta previa with short cervix (14.04% versus 5.69%); placenta previa with pre-eclampsia (13.45% versus 3.04%); hypothyroidism with short cervix (12.9% versus 2.03%); hypothyroidism with respiratory infections (12.9% versus 5.41%). In addition, FGR co-occurred with a short cervix in 10.43% of PTB cases compared to 3.74% of non-PTB cases (Table S6).

Several of these trends were preserved or further amplified in the SOL subset. Polyhydramnios with a short cervix was present in 63.64% of PTB cases versus just 8.00% in non-PTB cases. Likewise, perinatal sepsis and PROM co-occurred in 32.5% of PTB versus 16.08% of non-PTB. Other notable pairings included PROM with a short cervix (17.76% versus 2.46%) and placenta previa with a short cervix (18.4% versus 6.26%). The history of previous PTB also remained a frequently enriched factor when combined with perinatal sepsis and a short cervix (Table S7). These patterns suggested that a short cervix, infections, and abnormal amniotic fluid levels might play a role in triggering spontaneous PTB.

Although the CGI subset was limited by a smaller sample size (only 31 PTB cases), a few high-prevalence pairings were observed. Maternal complications dominated this group. Pre-eclampsia co-occurred with chorioamnionitis and placenta previa exclusively in PTB cases, and perinatal sepsis with PROM was present in 80.00% of PTB versus 10.00% of non-PTB. A previous history of PTB was observed in enriched pairings with pre-eclampsia and placenta previa (Table S8).

In the pairwise analysis performed among the SOL subset, PTB clusters were notably enriched for infection-related combinations (Fig. 2C, Table S9). All PTB cases in this group showed the presence of perinatal sepsis with FGR. Other significantly over-represented combinations included perinatal sepsis with parity, PROM, placenta previa; short cervix with PROM (5%), placenta previa, and parity; UTI with PROM and respiratory infections; and FGR with PROM. Additionally, a short cervix combined with biomass fuel exposure was more frequently observed in PTB clusters. In contrast, the CGI subset did not show any statistically significant associations after P-value adjustment, likely due to smaller sample sizes or weaker correlations. The gestational age for each variable is provided in Table S10.

Neonatal outcomes

The analysis revealed significantly higher NICU admission and early neonatal death (END) rates in PTB cases compared to non-PTB cases. The variables with the highest NICU admission rates among PTB were perinatal sepsis (100%), hypertension (76.92%), and pre-eclampsia (62.07%), with the lowest rates for anemia (25.63%) and placenta previa (28.37%), as shown in Table S10.

Association of PTB phenotypes with infant clinical outcomes

Using term deliveries as the reference group, we evaluated the association between infant clinical outcomes and PTB phenotypes (Table S11). Four conditions were significantly more prevalent in the PTB phenotype after multiple testing correction (P_adj < 0.01): respiratory distress, septic shock linked to perinatal sepsis, recurrent illness with anemia, and CNS disorders associated with hypothyroidism. In the SOL-PTB phenotype, CNS disorders also showed a strong association with hypothyroidism. No significant associations were observed in the CGI-PTB phenotype, likely due to limited sample size. Overall, these findings highlight perinatal sepsis in the PTB phenotype as a major risk factor for respiratory and septic complications in infants, while hypothyroidism consistently co-occurred with infant CNS disorders in both PTB and SOL-PTB phenotypes.

Discussion

Our study highlights the heterogeneity of the PTB phenotype (Fig. 3) and identifies key associated characteristics. PTB often occurs with multiple co-occurring conditions, including maternal conditions like pre-eclampsia, short cervix, perinatal sepsis, and a history of previous PTB, fetal factors such as polyhydramnios, oligohydramnios, and parturition-related complications like PROM. CGI cases were mainly related to maternal hypertensive disorders, while SOL cases were more linked to perinatal sepsis and PROM.

Fig. 3.

The figure illustrates the relationships and shared occurrences of various phenotypic variables among (A) PTB (left) and (B) non-PTB (right) cases. Each arc represents a distinct phenotypic variable, with connecting lines indicating co-occurrences with other variables. The thickness of the connecting lines and the proportional arc lengths reflect the frequency of co-occurrence across the dataset

Many of the features previously associated with PTB include PROM, which contributes to about one-third global preterm deliveries [15, 16]. PROM is also linked to the development of intrauterine infections [17], leading to serious complications such as perinatal sepsis. A study in China found that gestational hypertension and pre-eclampsia increase the risk of PTB [18]. Additionally, women with a cervical length shorter than 25 mm (10th percentile) at 24 weeks had a six-fold higher risk of spontaneous PTB before 35 weeks than those with lengths above 40 mm (75th percentile) [19]. Our findings emphasize the complex interplay between phenotypic features associated with PTB and variations in neonatal outcomes among PTB cases.

The clustering-based enrichment analysis revealed distinct phenotypic clusters related to PTB with varying co-occurring risk factors. In the SOL dataset, clusters linked PTB to maternal, placental, and parturition-related conditions like perinatal sepsis, short cervix, placenta previa, and PROM. Conversely, the CGI dataset identified maternal complications such as hypertension, respiratory infections, and hypothyroidism as key factors along with FGR. These findings emphasize the diverse causes of PTB and the need to consider different risk factors for both spontaneous and caregiver-induced cases.

Clustering analysis identified five key clinical feature combinations linked to PTB: perinatal sepsis with fetal growth restriction (FGR), parity, PROM, or placenta previa; and UTI with respiratory infections. These combinations increase PTB risk beyond individual factors, indicating heightened infection and obstetric vulnerability. Our findings align with previous studies [8, 9] and demonstrate the consistency of phenotypes across populations, particularly the co-occurrence of perinatal sepsis, FGR, and parity in multiple births.

This is the first study in LMIC settings, like India, to evaluate the phenotype of PTB beyond traditional classifications. It profiles PTB characteristics and maternal, fetal, and placental factors that could be targeted for preventive interventions if causal. Additionally, it enables the development of prediction models targeted towards the highly prevalent phenotypes, prioritizing those associated with adverse neonatal outcomes. Future steps include prospective validation and integration into clinical workflows to translate these findings into practice.

Limitations

The population diversity in this study represents only a subset of India’s broader demographic landscape and may not reflect populations in other countries. Existing literature shows that lower socioeconomic status, such as unemployment and lower educational attainment, is linked to an higher risk of preterm birth [20, 21]. In our cohort, a large proportion of unemployed participants with low levels of formal education, reflecting the population served by a public secondary-care hospital. This relative socioeconomic homogeneity limited our ability to assess the independent or modifying effects of socioeconomic factors on PTB risk and restricts the generalizability of our findings to more diverse populations.

Certain neonatal conditions, like septic shock, were defined in the cohort dataset using proxy variables rather than standard clinical criteria. For septic shock, this included the presence of systemic infection and the administration of intravenous fluids. While this approach allowed us to capture these events within the limits of our data, it may not fully align with standard diagnostic definitions and should be interpreted with caution.

A subset of data was used because participants were lost to follow-up or delivered at external sites. However, the risk of systematic loss-to-follow-up bias was minimal, as the baseline characteristics of participants included and excluded from the study were similar. Additionally, certain phenotypic features had relatively lower frequencies, which may have limited the statistical power to detect significant differences.

A key limitation of our clustering strategy is the exclusion of intermediate clusters (40–80% PTB, “grey zone”). While this enhances robustness and interpretability by minimizing misclassification, these phenotypically heterogeneous clusters may potentially represent transitional or early-risk PTB phenotypes. This grey zone is also contributed to by PTB cases with unknown causes. Furthermore, excluding grey-zone clusters may limit the generalizability of our findings to the full clinical spectrum of PTB. Future longitudinal studies should include these grey-zone cases to enhance individualized risk assessment. Additionally, NICU admissions included in this study may result from several factors unrelated to the phenotypic features being assessed. Lastly, the associated phenotypic features are associations useful for characterizing PTB and should not be interpreted as causal relationships.

The current study is based on the available data and utilizes unsupervised clustering to characterize PTB phenotypic heterogeneity. Predictive and ranking-based analyses may provide further insights into our findings. However, this would require larger sample sizes and independent validation in future studies. In addition, as analyses were based on complete cases, the possibility of bias due to non-random missingness cannot be excluded.

Conclusion

Our study highlights key maternal, fetal, and parturition features and their combinations, including placenta previa, perinatal sepsis, PROM, polyhydramnios, and medical disorders, which comprise a critical part of the PTB phenotype. Clinically, this detailed phenotyping can guide targeted interventions, enabling healthcare providers to offer appropriate prenatal care for high-risk pregnancies, ultimately improving maternal and neonatal outcomes. Expanding such studies with a more comprehensive dataset across diverse LMIC populations would further elucidate the PTB phenotype, providing a stronger foundation for the development of risk-stratification tools.

Abbreviations

SOL

Spontaneous onset of labor

PTB

Preterm birth

CGI

Caregiver-initiated birth

PROM

Pre-labor rupture of the membrane

NICU

Neonatal Intensive Care Unit

END

Early neonatal death

CNS

Central nervous system

FGR

Fetal growth restriction

LMIC

Low and middle-income countries

Authors’ contributions

RT, HS and SB conceived this study; VPG, Ayushi, DS and Aditi Sadhu preprocessed the data; VPG and Aditi Sadhu performed the exploratory data analysis; VPG, RJ and RT derived the phenotypic features; VPG, SBT, Aditi Sadhu and Ayushi performed data and statistical analyses; VPG, Ayushi, RT and HS verified the data and compiled the results; NW, Alka Singh, and UM, helped conduct and supervise the cohort study; VPG, SR, RT and HS interpreted the results; VPG, RT and HS wrote the first draft of the manuscript and VPG, SB, SR, RT and HS critically revised and edited subsequent manuscript drafts; all authors reviewed and approved the manuscript and had full access to the data.

Funding

The GARBH-Ini cohort study was funded by the Department of Biotechnology, Government of India (BT/PR9983/MED/97/194/2013). The ultrasound repository was partly supported by the Grand Challenges India-All Children Thriving Program, the Biotechnology Industry Research Assistance Council, and the Department of Biotechnology, Government of India (BIRAC/GCI/0115/03/14-ACT). The research reported in this publication was made possible by a grant (BT/kiData0394/06/18) from the Grand Challenges India at Biotechnology Industry Research Assistance Council (BIRAC), an operating division jointly supported by DBT-BMGF-BIRAC. An alum endowment from Prakash Arunachalam (BIO/18–19/304/ALUM/KARH) and Data Science and AI Consortium grant by the Ministry of Education, Government of India (SB/22–23/1256/CSETWO/008156) partly funded this study at the Centre for Integrative Biology and Systems Medicine, Wadhwani School of Data Science and AI, IIT Madras.

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request and after approval of the DBT Steering Committee of GARBH-Ini.

Declarations

Ethics approval and consent to participate

Full ethics approvals were obtained from the respective Institutional Ethics Committees/Institutional Review Boards of Translational Health Science and Technology Institute, District Civil Hospital (Gurugram), Safdarjung Hospital (New Delhi) (ETHICS/GHG/2014/1.43), and the Indian Institute of Technology Madras (IEC/2019–03/HS/01/07). Written informed consent was obtained from all study participants enrolled in the GARBH-Ini cohort. All methods were carried out in accordance with applicable guidelines, regulations, and in compliance with the 1964 Helsinki Declaration and its subsequent amendments.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.