From clusters to clinical rules: unsupervised machine learning identifies four newborn hearing phenotypes with bedside risk stratification Authors: Chen Liu, Jiali Zhang, Liu Yang Published: 2026 ABSTRACT ObjectiveTo identify risk patterns among high-risk newborns associated with hearing screening failure using unsupervised machine learning.Study designRetrospective cohort of 447 newborns. Partition around medoids clustering (Gower distance) was applied to demographic, perinatal, and diagnostic data. CART decision tree and diagnostic co-occurrence network analyses were performed to translate phenotypes into clinical rules and reveal risk factor synergy.ResultsFour distinct phenotypes emerged: Cluster 1 (28.4%, term jaundice) had the lowest failure (20.5%); Cluster 2 (25.5%, preterm males with respiratory morbidity) the highest (48.2%); Cluster 3 (30.6%, preterm females) intermediate (38.0%); Cluster 4 (15.4%, term respiratory/infectious) moderate (24.6%) (all P < 0.001). The decision tree generated simple bedside rules (e.g., males with weight < 1.795 kg → 83.3% failure). Co-occurrence network revealed a tightly connected core of respiratory disorders, infection, and preterm/low birth weight (Jaccard 0.31–0.46), while jaundice was isolated.ConclusionsUnsupervised clustering identified ~ ~ four neonatal phenotypes ~ ~ four distinct risk co-occurrence patterns in a high-risk neonatal unit cohort. Complementary decision tree and network analyses provided clinically actionable rules and exposed synergistic risk factor patterns that logistic regression could not capture. These findings support pattern-based risk stratification as a valuable complement to variable-centered methods.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12887-026-07203-9. Introduction Newborn hearing screening (NHS) is a critical public health intervention aimed at early detection of permanent childhood hearing impairment, enabling timely intervention and optimal language development [1, 2]. Universal NHS programs typically employ two complementary techniques: otoacoustic emissions (OAE), which assess outer hair cell function, and automated auditory brainstem response (AABR), which evaluates the entire auditory pathway up to the brainstem [3, 4]. Despite widespread implementation, a substantial proportion of newborns fail initial screening, necessitating diagnostic follow-up and imposing psychological and economic burdens on families and healthcare systems [5]. A recent systematic review reported that two-stage screening strategies yield referral rates ranging from 1.3% to 5.9% depending on the test combination used [6]. Several risk factors for NHS referral have been identified, including low birth weight, prematurity, hyperbilirubinemia, neonatal infections, and perinatal asphyxia [7, 8]. Premature infants are 50% more likely to develop hearing loss, with rates as high as 19% in extremely preterm infants (< 28 weeks) compared to < 1% in term infants [7]. However, these risk factors often co-occur, and traditional analyses based on single variables may not capture their complex interplay. Unsupervised machine learning approaches, such as cluster analysis, offer a data-driven means to identify distinct clinical phenotypes without prior assumptions, potentially revealing high-risk subgroups that transcend conventional classifications [9, 10]. Recent applications in neonatology have demonstrated the utility of clustering for identifying clinically meaningful phenotypes in extremely low birth weight infants with differential mortality and morbidity outcomes [9]. Such approaches may similarly enhance our understanding of hearing screening outcomes. In this study, we applied partition around medoids (PAM) clustering based on Gower distance to a cohort of newborns undergoing routine NHS. Our aim was to identify clinically meaningful subgroups defined by demographic, perinatal, and diagnostic features in a data-driven manner. Crucially, we sought to assess whether this clustering approach could uncover distinct patterns of risk factor co-occurrence that might be missed by traditional, variable-centered analyses (e.g., logistic regression). By comparing the insights derived from clustering with those from conventional methods, we aim to demonstrate the potential value of unsupervised learning as a complementary tool for risk stratification and hypothesis generation in newborn hearing screening research. Methods Study population and data collection This retrospective cohort study included all newborns admitted to the neonatal unit of a tertiary hospital in China between December 2024 and October 2025 who underwent routine hearing screening before discharge.This cohort represents a high-risk neonatal population; healthy term infants without any documented diagnoses were not included.Data were extracted from electronic medical records and included: sex, hospital ID (anonymized), birth date, gestational age (in weeks and days, e.g., “38 + 5”), delivery mode (Cesarean section or vaginal delivery), birth weight (kg), clinical diagnoses (free text in English), and results of OAE and AABR screening (pass or refer). Sex was assigned by the attending physician based on phenotypic examination at birth.Ethical approval for this retrospective study was obtained from the Institutional Review Board of the Maternal and Child Health Hospital of Hubei Province, Tongji Medical College (Approval No. 2026-143-01). The requirement for informed consent was waived by the IRB due to the retrospective nature of the study and the use of anonymized clinical data.Because this was a retrospective study based on routine electronic medical records, intrauterine growth restriction (IUGR) / fetal growth restriction was not available as a mandatory structured variable. We rechecked the extracted diagnostic fields and found no explicit documented diagnosis of IUGR or fetal growth restriction. Therefore, IUGR was not included as an analytic variable, and the absence of documented IUGR should not be interpreted as evidence of true absence in the clinical population. Data preprocessing Gestational age conversion Gestational age was converted to a numeric variable in weeks using the formula: weeks + days/7 (e.g., “38 + 5” to 38.714). For entries with only integer values (e.g., “39”), direct conversion was applied. Variable encoding Birth weight was used as a continuous variable (kg). Sex was coded as 1 for male and 0 for female; delivery mode was coded as 1 for Cesarean section and 0 for vaginal delivery. Screening outcomes were coded as binary variables: OAE_fail (1 = “not pass”, 0 = “pass”) and AABR_fail (1 = “not pass”, 0 = “pass”). Outcome definition The primary outcome was any screening failure, defined as a binary variable where 1 = failure on either OAE or AABR (i.e., OAE_fail = 1 or AABR_fail = 1), and 0 = pass on both tests. Secondary outcomes were OAE failure alone and AABR failure alone. This definition is used consistently throughout the manuscript. Diagnostic categorization Diagnostic text was preprocessed by splitting multiple diagnoses separated by Chinese commas (“, “) after conversion to standard commas. Based on clinical expertise and established risk factors for hearing loss [7, 8], each diagnosis was mapped to one or more of eight predefined categories. Based on clinical expertise and established risk factors for hearing loss, each diagnosis was mapped to one or more of eight predefined categories: respiratory disorders (e.g., respiratory distress syndrome, failure, asphyxia); jaundice (hyperbilirubinemia, ABO hemolytic disease); infections (sepsis, perinatal infection); metabolic disorders (hypoglycemia, maternal diabetes); prematurity/low birth weight (preterm, low birth weight, small for gestational age); cardiovascular disorders (arrhythmia); neurological disorders (brain injury); and other conditions (feeding difficulty, vomiting). For each infant, binary indicators were created for each category. Clustering analysis Distance calculation To handle mixed data types (continuous: gestational age, birth weight; categorical: sex, delivery mode, and eight binary diagnostic variables), we employed Gower distance, which computes dissimilarity as a weighted average of contributions from each variable [11]. Continuous variables were standardized to mean 0 and standard deviation 1 prior to distance calculation. The Gower distance matrix was calculated using the daisy function in R (version 4.2.1; package cluster). Clustering algorithm Partition around medoids (PAM) was chosen for its robustness to outliers and ability to use any dissimilarity matrix [12]. Unlike k-means, PAM identifies actual data points as cluster centers (medoids), enhancing interpretability. Determining optimal cluster number To determine the optimal number of clusters (K), we evaluated two metrics for K ranging from 2 to 10. Average silhouette width: measures how similar an object is to its own cluster compared to other clusters; values range from − 1 to 1, with higher values indicating better-defined clusters [13].Total within-cluster dissimilarity (elbow method): the sum of distances between points and their cluster medoids. Average silhouette width: measures how similar an object is to its own cluster compared to other clusters; values range from − 1 to 1, with higher values indicating better-defined clusters [13]. Average silhouette width: measures how similar an object is to its own cluster compared to other clusters; values range from − 1 to 1, with higher values indicating better-defined clusters [13]. Total within-cluster dissimilarity (elbow method): the sum of distances between points and their cluster medoids. Total within-cluster dissimilarity (elbow method): the sum of distances between points and their cluster medoids. The final K was selected based on a combination of statistical fit and clinical interpretability. Cluster stability assessment Cluster stability was evaluated via bootstrap resampling with 100 replicates. For each bootstrap sample (random sampling with replacement, same size as original dataset), PAM clustering was performed, and Jaccard similarities between original and bootstrapped cluster assignments were calculated [14]. Higher similarity (typically > 0.6) indicates acceptable stability. Characterization of clusters and association with hearing screening outcomes For each cluster, we summarized:Continuous variables (gestational age, birth weight) as mean ± standard deviationCategorical variables as percentages. Continuous variables (gestational age, birth weight) as mean ± standard deviation Continuous variables (gestational age, birth weight) as mean ± standard deviation Categorical variables as percentages. Categorical variables as percentages. To visualize cluster profiles, we created: A heatmap of scaled variable means across clusters.A radar chart displaying standardized values for key variables. A heatmap of scaled variable means across clusters. A heatmap of scaled variable means across clusters. A radar chart displaying standardized values for key variables. A radar chart displaying standardized values for key variables. The prevalence of OAE failure, AABR failure, and failure of either test was calculated for each cluster. Differences in failure rates across clusters were tested using chi-square tests. Odds ratios (OR) with 95% confidence intervals (CI) were calculated using Cluster 1 (lowest risk) as the reference group. For visualization, we performed t-distributed stochastic neighbor embedding (t-SNE) based on the Gower distance matrix to project clusters into two-dimensional space [15]. Statistical significance was set at P < 0.05. All analyses were performed in R. Comparison with traditional risk factor analysis To evaluate whether the clustering approach provided additional insights beyond conventional variable-centered methods, we performed a multivariable logistic regression analysis [16]. The outcome variable was “any screening failure” (failure of either OAE or AABR). All variables used in the clustering analysis were included as predictors: gestational age (continuous), birth weight (continuous), sex (binary), delivery mode (binary), and the eight binary diagnostic categories (respiratory disorders, jaundice, infections, metabolic disorders, prematurity/low birth weight, cardiovascular disorders, neurological disorders, and other conditions). Odds ratios (OR) with 95% confidence intervals (CI) were calculated for each predictor. The results of the logistic regression were then qualitatively compared with the cluster profiles to identify similarities and differences in the risk patterns revealed by the two analytical approaches. Specifically, we examined whether the clustering analysis identified subgroups characterized by the co-occurrence of multiple risk factors, a pattern that traditional regression, which assumes additive and independent effects of predictors, cannot directly capture. Results Study population characteristics Of 447 newborns (239 males, 53.5%; mean gestation 36.1 ± 3.2 weeks, range 25–42; mean birth weight 2.67 ± 0.73 kg, range 0.81–4.18), 93 (20.8%) failed OAE, 138 (30.9%) failed AABR, and 142 (31.8%) failed at least one test. These rates align with reported high-risk population norms (healthy: 2–4%; high-risk: 2–4%) [7, 17].Of these, 31 (6.9%) failed OAE only, 76 (17.0%) failed AABR only, and 62 (13.9%) failed both tests. No explicit diagnosis of IUGR or fetal growth restriction was documented in the extracted electronic medical records. Because IUGR was not captured as a structured variable, it was not included in the clustering analysis. Determination of optimal cluster number Silhouette analysis showed average width increasing from 0.18 at K = 2 to a peak of 0.41 at K = 9, with a local maximum of 0.35 at K = 4 (Figure S1). The elbow plot flattened after K = 4 (Figure S2). Based on statistical metrics and clinical interpretability, we selected K = 4. Cluster profiles The four clusters exhibited distinct clinical characteristics (Table 1; Fig. 1). Cluster 1 (n = 127, 28.4%) – “Term Jaundice Phenotype” Term infants (38.8 ± 1.0 weeks) with normal birth weight (3.21 ± 0.36 kg) and high jaundice prevalence (78.0%). Respiratory disorders absent. Lowest failure rates (OAE 4.7%, AABR 20.5%). Cluster 2 (n = 114, 25.5%) – “High-Risk Preterm Male Phenotype” Exclusively preterm males (mean 34.5 ± 2.6 weeks, range 25–42; the majority were late preterm [32–<37 weeks] according to WHO classification) with high respiratory (65.8%) and prematurity/LBW diagnoses (97.4%), infections common (20.2%). Highest failure rates (OAE 30.7%, AABR 47.4%, any 48.2%). Cluster 3 (n = 137, 30.6%) – “Intermediate-Risk Preterm Female Phenotype” Exclusively preterm females (35.1 ± 2.6 weeks, 2.18 ± 0.53 kg) with frequent respiratory (65.0%), prematurity/LBW (86.9%), and infections (22.6%). Intermediate failure rates (OAE 26.3%, AABR 35.8%, any 38.0%). Cluster 4 (n = 69, 15.4%) – “Respiratory-Infection Phenotype” Term infants (38.8 ± 1.3 weeks, 3.23 ± 0.35 kg) with high respiratory (84.1%) and infections (63.8%). Moderate failure rates (OAE 15.9%, AABR 23.2%, any 24.6%). Table 1Clinical characteristics and hearing screening outcomes by clusterCharacteristicCluster 1 (n = 127)Cluster 2 (n = 114)Cluster 3 (n = 137)Cluster 4 (n = 69)Gestational age (weeks)38.8 ± 1.034.5 ± 2.635.1 ± 2.638.8 ± 1.3Birth weight (kg)3.21 ± 0.362.14 ± 0.522.18 ± 0.533.23 ± 0.35Male (%)59.8100069.6Respiratory disorders (%)065.865.084.1Jaundice (%)78.001.50Infections (%)7.920.222.663.8Metabolic disorders (%)12.69.610.25.8Prematurity/LBW (%)2.497.486.91.4Cardiovascular (%)1.600.71.4Neurological (%)0001.4Other conditions (%)5.503.65.8OAE failure (%)4.730.726.315.9AABR failure (%)20.547.435.823.2Any failure (%)20.548.238.024.6Data are presented as mean ± SD or %. LBW Low birth weight Table 1Clinical characteristics and hearing screening outcomes by clusterCharacteristicCluster 1 (n = 127)Cluster 2 (n = 114)Cluster 3 (n = 137)Cluster 4 (n = 69)Gestational age (weeks)38.8 ± 1.034.5 ± 2.635.1 ± 2.638.8 ± 1.3Birth weight (kg)3.21 ± 0.362.14 ± 0.522.18 ± 0.533.23 ± 0.35Male (%)59.8100069.6Respiratory disorders (%)065.865.084.1Jaundice (%)78.001.50Infections (%)7.920.222.663.8Metabolic disorders (%)12.69.610.25.8Prematurity/LBW (%)2.497.486.91.4Cardiovascular (%)1.600.71.4Neurological (%)0001.4Other conditions (%)5.503.65.8OAE failure (%)4.730.726.315.9AABR failure (%)20.547.435.823.2Any failure (%)20.548.238.024.6Data are presented as mean ± SD or %. LBW Low birth weight Clinical characteristics and hearing screening outcomes by cluster Data are presented as mean ± SD or %. LBW Low birth weight Fig. 1Radar chart of clinical characteristics for the four clusters. Radar chart displaying the standardized mean values of key clinical variables for each of the four identified clusters. The variables include gestational age (gestation_week_num_scaled), birth weight (weight_scaled), and the percentages of male infants (gender_code_pct), Cesarean delivery (delivery_code_pct), and each diagnostic category: respiratory disorders (dx_resp_pct), jaundice (dx_jaundice_pct), infections (dx_infect_pct), metabolic disorders (dx_metab_pct), prematurity/low birth weight (dx_preterm_lbw_pct), cardiovascular disorders (dx_cardio_pct), neurological disorders (dx_neuro_pct), and other conditions (dx_other_pct). Each axis represents one variable scaled to a common range to facilitate comparison across clusters Radar chart of clinical characteristics for the four clusters. Radar chart displaying the standardized mean values of key clinical variables for each of the four identified clusters. The variables include gestational age (gestation_week_num_scaled), birth weight (weight_scaled), and the percentages of male infants (gender_code_pct), Cesarean delivery (delivery_code_pct), and each diagnostic category: respiratory disorders (dx_resp_pct), jaundice (dx_jaundice_pct), infections (dx_infect_pct), metabolic disorders (dx_metab_pct), prematurity/low birth weight (dx_preterm_lbw_pct), cardiovascular disorders (dx_cardio_pct), neurological disorders (dx_neuro_pct), and other conditions (dx_other_pct). Each axis represents one variable scaled to a common range to facilitate comparison across clusters Association with hearing screening outcomes Failure rates differed significantly across clusters for both OAE (χ² = 35.2, P < 0.001) and AABR (χ² = 29.1, P < 0.001) (see Table 1 and Supplementary Figure S5). Using Cluster 1 as the reference group:Cluster 2 had the highest risk of any screening failure (OR 3.6, 95% CI 2.2–5.9, P < 0.001).Cluster 3 also showed significantly elevated risk (OR 2.4, 95% CI 1.5–3.8, P < 0.001).Cluster 4 did not differ significantly from Cluster 1 (OR 1.3, 95% CI 0.7–2.3, P = 0.45). Cluster 2 had the highest risk of any screening failure (OR 3.6, 95% CI 2.2–5.9, P < 0.001). Cluster 2 had the highest risk of any screening failure (OR 3.6, 95% CI 2.2–5.9, P < 0.001). Cluster 3 also showed significantly elevated risk (OR 2.4, 95% CI 1.5–3.8, P < 0.001). Cluster 3 also showed significantly elevated risk (OR 2.4, 95% CI 1.5–3.8, P < 0.001). Cluster 4 did not differ significantly from Cluster 1 (OR 1.3, 95% CI 0.7–2.3, P = 0.45). Cluster 4 did not differ significantly from Cluster 1 (OR 1.3, 95% CI 0.7–2.3, P = 0.45). The t-SNE projection based on Gower distance demonstrated reasonable separation between clusters, particularly for Cluster 2 (Figure S3), supporting the distinctiveness of the identified phenotypes. Complementary insights from clustering compared to logistic regression To evaluate whether clustering provides added value beyond conventional variable-centered methods, we performed multivariable logistic regression with the same variables used in clustering (see Methods). The logistic regression model identified male sex (OR 2.1, 95% CI 1.4–3.2, P < 0.001), lower gestational age (OR 0.89 per week, 95% CI 0.82–0.96, P = 0.003), and presence of respiratory disorders (OR 1.9, 95% CI 1.2–3.0, P = 0.006) as independent risk factors for any screening failure. Other variables, including birth weight and diagnostic categories such as infection or jaundice, did not reach statistical significance. The two methods answer different but complementary questions. Logistic regression identifies population-average independent risk factors (which factors matter on average? ). Clustering identifies multidimensional risk profiles (phenotypes) (how do these factors co-occur in specific subgroups? ). For example, Cluster 2 was characterized by the co-occurrence of male sex, preterm gestation (mean 34.5 weeks), low birth weight (mean 2.14 kg), and high prevalence of respiratory disorders (65.8%) and infections (20.2%). The failure rate in this cluster was 48.2%, which is substantially higher than the 41.2% predicted by a logistic model that included the same three main effects (male, GA = 34.5 weeks, respiratory disorder) without interactions. This suggests that the co-occurrence of multiple risk factors within a specific subgroup may capture synergistic or cumulative effects that are not directly evident from main‑effects regression. Thus, clustering serves as a hypothesis-generating tool that identifies candidate high-risk subgroups (e.g., “preterm males with respiratory morbidity”), whose risk profile can then be validated in subsequent studies. We do not claim superiority of one method over the other; rather, we argue that they are complementary. A side‑by‑side comparison of the two approaches is provided in Supplementary Table S1. Additionally, we note that the outcome (32% failure) is moderately imbalanced, which logistic regression can handle but may benefit from weighting or resampling – this is acknowledged as a consideration for future work. Decision tree risk stratification To translate the clustering findings into clinically actionable rules, we constructed a classification and regression tree (CART) using all predictor variables (Supplementary Table S2). The pruned tree identified seven terminal nodes, with birth weight (2.165 kg) as the primary split. Infants weighing ≥ 2.165 kg (Node 2, n = 330) had a failure rate of 26.4%. Among those with lower birth weight, sex further stratified risk: males with weight < 1.795 kg exhibited the highest failure rate (83.3%, Node 15), whereas females with gestational age ≥ 33.6 weeks had a relatively lower rate (32.3%, Node 12). These decision rules align closely with the four clusters identified by PAM and provide an intuitive framework for bedside risk assessment (Fig. 2). Fig. 2Decision tree for newborn hearing screening failure. The tree was pruned using cross-validation (cp = 0.01667). Each node displays the decision rule, number of failures/total cases, and the predicted outcome (0 = pass, 1 = fail). Terminal nodes correspond to the risk strata described in Supplementary Table S2 Decision tree for newborn hearing screening failure. The tree was pruned using cross-validation (cp = 0.01667). Each node displays the decision rule, number of failures/total cases, and the predicted outcome (0 = pass, 1 = fail). Terminal nodes correspond to the risk strata described in Supplementary Table S2 Diagnosis co‑occurrence network To explore synergistic relationships among diagnostic categories, we constructed a co‑occurrence network based on Jaccard similarity (Supplementary Table S3). Three diagnoses—respiratory disorders, infection, and preterm/low birth weight—formed a tightly connected core subnetwork (similarities 0.31–0.46). In contrast, jaundice showed negligible co‑occurrence with all other categories (similarity ≤ 0.05), consistent with its isolation in the low‑risk Cluster 1. Degree centrality identified respiratory disorders and preterm/low birth weight as hub nodes, highlighting their pivotal roles in shaping high‑risk phenotypes (Fig. 3). Fig. 3Diagnosis co‑occurrence network. Nodes represent diagnostic categories, sized by degree centrality. Edges represent Jaccard similarity > 0.1, with thickness proportional to similarity. Full similarity values are provided in Supplementary Table S3 Diagnosis co‑occurrence network. Nodes represent diagnostic categories, sized by degree centrality. Edges represent Jaccard similarity > 0.1, with thickness proportional to similarity. Full similarity values are provided in Supplementary Table S3 Cluster stability Bootstrap analysis revealed good stability for the four-cluster solution, with median Jaccard similarities of 0.78 (IQR 0.71–0.84) for Cluster 1, 0.81 (IQR 0.74–0.87) for Cluster 2, 0.76 (IQR 0.68–0.82) for Cluster 3, and 0.73 (IQR 0.65–0.80) for Cluster 4 (Figure S4). All values exceeded the recommended threshold of 0.6 for acceptable stability [18], indicating that the clusters were robust to sampling variability. Quantitative evidence that clustering captures co‑occurrence patterns beyond demographic splits To address whether the clustering solution merely reflects simple demographic splits (e.g., sex or gestational age), we performed five complementary analyses summarized in Supplementary Table S4. The key finding is that Cluster 2 was not defined by male sex alone, but by the combined pattern of male sex, preterm birth, low birth weight, respiratory morbidity, and infection. Direct comparison between Cluster 2 (all male) and Cluster 3 (all female) showed nearly identical gestational age (34.5 ± 2.6 vs. 35.1 ± 2.6 weeks, p = 0.07) and respiratory morbidity (65.8% vs. 65.0%, p = 0.89), yet Cluster 2 had a significantly higher failure rate (48.2% vs. 38.0%; OR = 1.52, 95%CI 1.02–2.27, p = 0.038). This indicates that the excess risk in Cluster 2 is not explained by gestational age or respiratory morbidity alone. In a homogeneous subgroup restricted to late‑preterm infants (34–37 weeks) with respiratory disorders, males had a higher failure rate than females (72.1% vs. 40.0%; OR = 3.68, 95%CI 1.33–10.18, p = 0.009), confirming that sex acts as an effect modifier within a high‑risk clinical context. Multivariable logistic regression in preterm infants further showed that male sex remained independently associated with failure after adjustment for gestational age, birth weight, and respiratory disease (OR = 2.24, 95%CI 1.31–3.82, p = 0.003). Importantly, male predominance alone was insufficient to explain the clustering result: cross‑cluster comparison of male‑predominant clusters revealed that Cluster 1 (59.8% male) and Cluster 4 (69.6% male) had low‑to‑moderate failure rates (20.5% and 24.6%, respectively), whereas the all‑male Cluster 2 had a failure rate of 48.2%. The diagnostic co‑occurrence network further supported this pattern, showing that respiratory disorders, infection, and preterm/low birth weight formed a tightly connected core (Jaccard similarities 0.31–0.46), while jaundice remained isolated. Collectively, these analyses demonstrate that clustering captured a high-risk co-occurrence pattern (male + preterm + low birth weight + respiratory morbidity + infection) rather than a simple sex- or gestational-age-based split. This pattern would not be automatically evident from univariate or main-effects regression. Discussion Using unsupervised clustering, we identified four neonatal phenotypes with distinct perinatal and diagnostic profiles. Clusters 2 (exclusively male) and 3 (exclusively female) comprised preterm, low-birth-weight infants with high respiratory morbidity but differed in failure rates—Cluster 2 had the highest risk, Cluster 3 intermediate. Cluster 4 (term, respiratory/infectious) showed moderate risk, while Cluster 1 (term, jaundice) had the lowest. Prematurity, low birth weight, and hearing loss risk Our findings confirm the association between prematurity, low birth weight, and hearing screening failure [7, 8]. Preterm hearing impairment rates are inversely related to gestational age and birth weight (19% at < 28 weeks, 18% at < 1 kg). In our cohort, Clusters 2 and 3 (34–35 weeks, 2.1–2.2 kg) had failure rates of 38–48%, far exceeding the < 1% in term infants [7]. This discrepancy likely reflects cumulative risk factors (ototoxic medication, ventilation, illness) common in preterm populations [8, 19]. Methodological considerations: the value of pattern-based discovery A key contribution of this study lies in its demonstration of how unsupervised clustering can complement traditional analytical approaches. Logistic regression, the dominant paradigm in clinical risk factor research, assumes that predictors contribute independently and additively to the outcome. This variable-centered approach is valuable for identifying “what” factors matter. However, as our comparative analysis illustrates, it may overlook clinically important subgroups defined by “how” these factors cluster together. In our cohort, Cluster 2 represents precisely such a subgroup. While regression confirmed that male sex, prematurity, and respiratory disorders are each independent risk factors, it could not reveal that the co-occurrence of these three factors in a single subgroup (25.5% of the cohort) would be associated with a failure rate (48.2%) substantially higher than the sum of their individual effects. This pattern suggests potential synergistic or cumulative effects that warrant further investigation. From this perspective, clustering serves as a powerful hypothesis-generating tool: it identifies candidate subgroups (e.g., “preterm males with respiratory morbidity”) whose risk profile can then be validated and explored in subsequent studies using confirmatory methods. Recent studies across various medical domains have demonstrated similar utility of unsupervised learning for enhancing risk stratification [20, 21]. We argue that integrating unsupervised learning with traditional regression offers a more complete understanding of complex clinical phenomena than either approach alone. Sex-specific differences in vulnerability A striking finding was the complete sex segregation between Clusters 2 (male) and 3 (female), despite similar gestational age, birth weight, and respiratory morbidity. Males had significantly higher failure rates, suggesting male sex confers additional vulnerability in preterm infants. This aligns with reports of higher hearing screening failure in male neonates and that 78% of children with hearing deficits are males [7, 22]. The mechanisms underlying sex-specific auditory differences remain unclear but may involve multiple factors. Males born preterm exhibit lower cognitive scores and greater motor impairment, linked to differences in white matter maturation [23]. Males show lower superior white matter fractional anisotropy, and sex moderates associations between brain injury, pain exposure, and neurodevelopmental outcomes [23]. In extremely preterm infants, males are 20% less likely to survive without moderate/severe neurodevelopmental impairment [24]. These differences may extend to the auditory system, explaining higher hearing screening failure rates in males. Respiratory disorders and infection as risk factors Cluster 4, despite term gestation and normal birth weight, was dominated by respiratory and infectious diagnoses. Its moderate failure rate (24.6%) suggests that acute perinatal illnesses can impact auditory function even without prematurity. Potential mechanisms include hypoxia-ischemia, systemic inflammation, and ototoxic medication exposure during intensive care [8, 25]. A case-control study of preterm infants (< 32 weeks) found that cumulative exposure to ototoxic medications (gentamicin, vancomycin, furosemide) and physiological risk factors (acidosis, elevated bilirubin) independently increased hearing loss risk [8]. Each day of medication exposure or physiological risk was associated with a 23% and 15% increase in odds, respectively [8].This cumulative risk model may explain why Cluster 4, with its high burden of respiratory illness requiring intensive care, has elevated failure rates despite term birth. Jaundice and hearing outcomes: a nuanced relationship The absence of jaundice in high-risk clusters (2–4) is notable, as hyperbilirubinemia is a classic risk factor for auditory neuropathy [26, 27]. Severe cases can cause present OAE but absent AABR [28]. In our cohort, jaundice was predominantly in Cluster 1 (best outcomes), likely reflecting mostly mild/physiological jaundice effectively managed; severe cases requiring exchange transfusion were rare, explaining the lack of association with screening failure. This aligns with evidence that bilirubin-induced neurological dysfunction depends on magnitude, duration, and individual susceptibility [27, 29]. Clinical implications These findings have several clinical implications. First, multidimensional risk stratification may outperform single-factor approaches. Cluster 2 (25.5% of the cohort) accounted for a disproportionate share of failures and may benefit from enhanced surveillance, prioritization of AABR screening, and closer follow-up. Second, sex-specific differences suggest that risk models should treat sex as an effect modifier, not a simple covariate. Third, the elevated risk in Cluster 4 indicates that term infants with significant respiratory or infectious morbidity should not be considered low-risk based on gestational age alone. Our findings also support recommendations for two-stage screening protocols that include AABR for high-risk populations [6]. While OAE alone may be sufficient for low-risk term infants (Cluster 1), the high rate of AABR failure in Clusters 2–4 (23–47%) justifies the use of AABR in neonatal intensive care unit populations, as recommended by the Joint Committee on Infant Hearing. Complementary value of decision tree and co‑occurrence network To further translate the clustering findings into clinically actionable tools, we applied CART decision tree analysis [30] and a diagnostic co‑occurrence network [31]. The decision tree generated simple “if‑then” rules based solely on birth weight, gestational age, sex, and respiratory status (Supplementary Table S2). These rules closely mirror the four PAM clusters and enable bedside risk stratification without computational support. Such simplification is particularly valuable in neonatal screening programs where resource constraints often limit complex scoring systems. The co‑occurrence network (Supplementary Table S3) revealed that three diagnostic categories—respiratory disorders, infection, and preterm/low birth weight—form a tightly connected core (Jaccard similarities 0.31–0.46), whereas jaundice remains isolated. This structural insight aligns with the clustering results: the high‑risk phenotypes (Clusters 2 and 3) are characterized by the co‑presence of these core diagnoses, while the low‑risk phenotype (Cluster 1) is dominated by isolated jaundice. Importantly, such co‑occurrence patterns are not captured by standard logistic regression but are naturally exposed by unsupervised learning and its complementary methods. Together, these analyses provide a multidimensional view of neonatal hearing risk and offer practical tools for tailored screening strategies. Comparison with previous clustering studies Our use of unsupervised clustering extends prior neonatology work [9]. Matsushita et al. identified six phenotypes in extremely low birth weight infants with differential mortality, showing clustering reveals subgroups missed by traditional classifications [9]. Similarly, we identified a highest-risk phenotype (Cluster 2: male sex, prematurity, low birth weight, respiratory morbidity)—a combination difficult to capture through conventional analysis. Strengths and limitations Strengths of this study include: (1) the use of a data-driven clustering approach that captures multidimensional interactions without prior assumptions; (2) the inclusion of detailed diagnostic information derived from clinical records; (3) rigorous cluster validation through silhouette analysis, elbow method, and bootstrap stability assessment; (4) the identification of clinically interpretable phenotypes with clear risk stratification; and (5) the novel finding of complete sex segregation in two otherwise similar preterm clusters. Limitations should be acknowledged. First, the single-center design and the lack of healthy term controls limit generalizability; our findings are derived from a high-risk neonatal unit cohort and require external validation in population-based settings. Second, our cohort did not include extremely preterm infants (< 28 weeks). In addition, IUGR/fetal growth restriction was not available as a reliably documented structured variable in the electronic medical records. Although no explicit IUGR diagnosis was identified in the extracted diagnostic fields, this should be interpreted as lack of documentation rather than definitive absence in the clinical population. Therefore, the potential contribution of IUGR to hearing screening failure could not be evaluated, which limits the generalizability of the proposed bedside risk stratification, especially for growth‑restricted neonates. Third, we used raw birth weight in the clustering model for clinical interpretability; future studies may explore gestational age- and sex-specific z-scores. Additionally, the class distribution of the primary outcome (32% failure) is moderately imbalanced, which logistic regression may be sensitive to; we have noted this in Sect. Complementary insights from clustering compared to logistic regression and suggest weighting or resampling in future work. Finally, lack of long-term audiological follow-up precluded confirmation of permanent hearing loss. Future directions Future research should validate these phenotypes in independent, multicenter cohorts with larger sample sizes. Longitudinal follow-up with diagnostic audiometry is needed to determine whether screening failures translate to confirmed permanent hearing loss and whether phenotypes predict long-term auditory and language outcomes. Integration of additional variables, such as bilirubin levels, medication exposure duration, mechanical ventilation days, and genetic factors (e.g., the m.1555 A > G mutation associated with aminoglycoside ototoxicity [8]), could refine the phenotypes and elucidate underlying mechanisms. Prospective studies could evaluate whether phenotype-based risk stratification improves the efficiency of follow-up programs and reduces loss to follow-up, which remains a significant challenge in NHS programs [6]. Conclusion Unsupervised clustering identified four neonatal phenotypes with distinct hearing screening failure risks. The highest-risk group comprised preterm, low-birth-weight males with respiratory morbidity; a female counterpart showed intermediate risk. Term infants with respiratory/infectious disease also had elevated failure rates. Importantly, a comparative analysis with logistic regression demonstrated that clustering revealed a pattern of risk factor co-occurrence—particularly in the highest-risk subgroup—that traditional variable-centered methods could not capture. These findings support the value of multidimensional, pattern-based risk stratification as a complement to conventional approaches and warrant validation in diverse populations with longitudinal outcomes. Supplementary information Supplementary Material 1. Supplementary Material 1.