Development of a non-invasive diagnostic model for severe Retinopathy of Prematurity integrating clinical and platelet data

Abstract

To construct a non-invasive and convenient early diagnostic model by integrating multidimensional clinical data and platelet (PLT) indices, and to explore the predictive value of PLT for severe Retinopathy of Prematurity (ROP). A study included premature infants admitted to our hospital from January 2020 to September 2025. According to the results of fundus screening, subjects were divided into the ROP group (n = 190) and the normal control group (n = 142). The ROP group was further categorized into mild (n = 110) and severe (n = 15) subgroups based on treatment requirements, which included platelet data corresponding to a postmenstrual age (PMA) of 30 weeks. Clinical data on parental factors, neonatal factors, and treatment factors were collected, along with PLT results from birth to PMA of 40 weeks. Lasso regression was used to select predictive variables, and a nomogram was constructed using multivariate logistic regression, with the model’s discrimination and calibration verified. Lasso regression identified gestational age, in vitro fertilization, maternal age, and PLT as core predictive factors. The Area Under the Curve (AUC) of the Nomogram in the training and validation sets was 0.80 (95%CI: 0.74–0.85) and 0.80 (95%CI: 0.71–0.89) respectively. The PLT levels at PMA of 30 weeks in the severe ROP group were significantly lower than those in the mild group (160 × 10^9/L vs. 254 × 10^9/L, p = 0.048), with the AUC of the Nomogram based on PLT combined with clinical indicators reaching 0.86 (95%CI: 0.76–0.96) for severe ROP. The ROP prediction model in this study can assist in the non-invasive identification of high-risk infants for ROP, particularly demonstrating high predictive efficacy for severe ROP.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-28293-y.

Introduction

Retinopathy of Prematurity (ROP) is a prevalent blinding eye disease affecting premature and low birth weight infants, fundamentally associated with abnormal retinal vascular development due to prematurity. Globally, advancements in perinatal medicine and neonatal rescue technologies have significantly increased the survival rate of premature infants; however, the incidence of ROP has also risen^1–3. If not addressed promptly, ROP may progress to retinal detachment, globe atrophy, and even permanent blindness, severely impacting the quality of life for affected children and imposing a socioeconomic burden on their families⁴.

Existing screening methods, such as indirect ophthalmoscopy and the Retcam fundus imaging system, encounter challenges, including operational complexity, the need for pupillary dilation and sedation, and a heavy reliance on medical resources, which hinder their widespread adoption in remote areas. Furthermore, repeated examinations may increase stress risk for infants⁵. Therefore, there is an urgent need for a rapid and efficient screening method in clinical practice. Recent studies have indicated that platelet (PLT) parameters correlate with abnormal vascular proliferation in ROP; however, these studies have primarily focused on the simple correlation between platelet parameters and ROP without considering other key clinical variables^6–9.

In light of this background, the present study aims to develop an efficient, non-invasive tool for predicting ROP and risk stratification by integrating platelet parameters with multidimensional clinical indicators. This approach seeks to optimize the allocation of screening resources and provide evidence-based support for individualized prevention strategies, ultimately reducing the blindness rate associated with ROP and improving the quality of life for premature infants.

Methods

Patients

The research flowchart is depicted in Fig. 1. We conducted a retrospective analysis of data from 1,644 infants admitted to the neonatal department of Children’s Hospital Affiliated to Shandong University, between January 1, 2020, and September 1, 2025, who were free of retinopathy of prematurity (ROP) at admission. A total of 902 infants were excluded due to being over 8 days old at admitted to the hospital, 297 infants were excluded for failing to complete follow-up by 40 weeks of postmenstrual age (PMA), 33 infants were excluded for not meeting the ROP screening criteria, 24 infants were excluded due to incomplete information, 15 infants were excluded due to blood disorders, and 41 infants were excluded due to parental abandonment or transfer to other hospitals. Ultimately, 332 infants with complete data were identified as eligible for establishing an early ROP diagnostic model.

Fig. 1.

Flowchart.*:children with missing PMA = 30w PLT were excluded.

All infants underwent ROP screening upon admission, adhering to international guidelines, which included a birth weight of less than 1,500 g, a gestational age of less than 30 weeks, or ROP risk factors identified by pediatricians or neonatologists. Based on the results of the fundus screening, infants were categorized into two groups: (1) those diagnosed with ROP before reaching 40 weeks of postmenstrual age (PMA), classified as the ROP group; and (2) those who did not develop ROP by 40 weeks of PMA, classified as the non-ROP group.

This study was conducted in accordance with the Declaration of Helsinki and received approval from the Research Ethics Committee of Children’s Hospital Affiliated to Shandong University. (Ethical approval number: SDFE-IRB/P-2024046). This study was a retrospective study, the use of anonymized information data for research complied with relevant regulations and ethical principles, Informed consent was waived by the Ethical Board of Children’s Hospital Affiliated to Shandong University.

Collection of information

Twenty-four clinical features were extracted from the electronic medical records of each infant. These features encompass five maternal factors, twelve neonatal factors and seven treatment factors, as detailed in Table 1. The collection of platelet data and their parameters is continuous, commencing from the first admission. Data are collected weekly according to the PMA until the conclusion of the follow-up period.

Table 1.

Demographic characteristics.

Variables Names, median(Q1-Q3)	non-ROP N = 142(%)	ROP N = 190(%)	P value
Parental factors
Father’s age, y	32.00 (28.00, 34.00)	33.00 (30.00, 35.00)	0.023
Mother’s Age, y	31.00 (27.00, 33.00)	32.00 (28.00, 35.00)	0.117
Maternal comorbidities	93 (65.49)	106 (55.79)	0.074
Maternal medication	89 (62.68)	123 (64.74)	0.699
IVF	8 (5.63)	26 (13.68)	0.017
Fetal factors
Age, d	1.50 (1.00, 4.00)	1.00 (1.00, 3.00)	0.041
Sex (boys)	92 (64.79)	124 (65.26)	0.929
Gestational age, wk	32.00 (30.29, 34.00)	29.07 (27.57, 31.00)	< 0.001
Cesarean delivery	101 (71.13)	97 (51.05)	< 0.001
5 min Apagr,	8.00 (7.00, 9.00)	7.00 (6.00, 9.00)	0.002
Birth Weight, g	1505.00 (1300.00, 1797.50)	1247.50 (1052.50, 1497.50)	< 0.001
Asphyxiation	59 (41.55)	116 (61.05)	< 0.001
Intrauterine distress	32 (22.54)	39 (20.53)	0.659
PLT, *109/l	179.50 (138.25, 247.75)	198.00 (142.00, 244.00)	0.428
MPV, fl.	10.10 (9.50, 10.80)	10.00 (9.40, 10.67)	0.587
PDW, %	16.40 (13.45, 16.98)	16.60 (15.40, 17.00)	0.109
P-LCR, %	26.85 (22.35, 31.67)	26.40 (22.00, 30.28)	0.734
Therapeutic factors
Oxygen usage	130 (91.55)	186 (97.89)	0.008
Initial exposure to MV	74 (52.11)	141 (74.21)	< 0.001
Duration of MV, d	2.00 (0.00, 5.00)	5.00 (1.00, 18.00)	< 0.001
Duration of non-invasive ventilation, d	16.00 (6.00, 29.75)	30.00 (18.25, 40.00)	< 0.001
Duration of high flow inhalation, d	3.00 (0.00, 8.00)	6.00 (2.00, 11.75)	< 0.001
Blood transfusion	91 (64.08)	155 (81.58)	< 0.001
Surgery	30 (21.13)	43 (22.63)	0.743

Prediction model development

We conducted Lasso regression on the aforementioned 24 variables to perform variable selection through the Least Absolute Shrinkage and Selection Operator. The variables identified by the Lasso regression were subsequently analyzed using logistic regression. Finally, we assessed the clinical applicability and simplicity of the model to develop a nomogram. The discriminative ability of the nomogram was assessed through Receiver Operating Characteristic (ROC) analysis and by calculating the Area Under the Curve (AUC). Additionally, calibration curves and the Hosmer-Lemeshow test were utilized to mitigate overfitting bias, and the goodness of fit was evaluated by comparing the actual probabilities with those predicted by our nomogram. To estimate the net benefit of the nomogram model at different threshold probabilities, Decision Curve Analysis (DCA) was conducted.

This study constructed two nomograms. The first nomogram aims to early differentiate between non-ROP and ROP children.The dataset was randomly divided into a training set and a validation set in a 7:3 ratio. The training set consisted of 98 non-ROP children and 134 ROP children, while the validation set comprised 44 non-ROP children and 56 ROP children. The training set was utilized to establish models using Lasso and logistic regression, while the validation set was employed to validate the final nomogram model. The second nomogram is designed to distinguish between mild and severe cases of ROP at an early stage. The sample group for this model includes both ROP cases requiring ranibizumab treatment and those with platelet data at 30 weeks of PMA. In total, there were 110 cases of mild ROP and 15 cases of severe ROP (Fig. 1).

Statistical method

Statistical analyses were conducted using SPSS version 17.0, with continuous data expressed as the median (25th − 75th percentile). Intergroup comparisons were performed using the Wilcoxon rank-sum test. Qualitative data were presented as counts (n/n), and intergroup comparisons were conducted using the chi-square test. A P-value of less than 0.05 was considered statistically significant.

Lasso regression analysis was performed using Python software (version 4.3.0) with 10-fold cross-validation to screen for predictive variables. Following the results of the Lasso regression, logistic regression analysis was employed to identify the optimal predictive features for ROP. Subsequently, a nomogram, calibration curve, receiver operating characteristic (ROC) curve, and decision curve analysis (DCA) were utilized to validate the model’s calibration, discriminative ability, and clinical effectiveness.

Results

Description of the patients

In this study, a total of 332 cases were included, with 190 classified into the ROP group and 142 into the non-ROP group based on the results of fundus screening. The analysis revealed statistically significant differences between the groups concerning parental factors (father’s age, in-vitro fertilization (IVF)), fetal factors (Age, gestational age, cesarean delivery, 5 min Apgar, birth weight, asphyxiation), and treatment factors (oxygen usage, initial exposure to mechanical ventilation (MV), duration of MV, duration of non-invasive ventilation, duration of high flow inhalation, blood transfusion) (P < 0.05). Other clinical features and laboratory indicators did not show statistically significant differences (P > 0.05) (Table 1).

Establishment of the early prediction model for ROP

By conducting Lasso regression analysis on all statistically significant features with ROP as the outcome indicator, five optimal predictive factors were identified: gestational age, mother’s age, duration of MV, IVF, and postnatal asphyxia (Fig. 2). These five predictive factors were subsequently employed in a logistic regression analysis to establish a nomogram model. It was observed that duration of MV and postnatal asphyxia contributed minimally to the model. To enhance model simplicity, only three predictive factors—gestational age, mother’s age, and IVF—were retained in the final model (Fig. 3). The results are as shown in Table 2: gestational age [Odds Ratio (OR): 0.64, 95% Confidence Interval (CI): 0.56–0.73], mother’s age (OR: 1.04, 95% CI: 0.98–1.11), and IVF (OR: 1.80, 95% CI: 0.65–5.03). This model provides a score for each variable, and the total score is derived by summing these individual scores. This total score corresponds to the length of a line segment, facilitating the estimation of the probability of ROP occurrence. The chart illustrates that factors such as younger gestational age, older mother’s age, and IVF yield higher scores, thereby increasing the risk of ROP.

Fig. 2.

Screening of 24 variables based on Lasso regression. A: The variation characteristics of the coefficient of variables; B: the selection process of the optimum value of the parameter log(l) in the Lasso regression model.

Fig. 3.

A: Nomogram model to predict the incidence of ROP. IVF: 0(No), 1(Yes); B: The ROC-AUC of the training and validation groups. C: The calibration curves for the training and validation groups.

Table 2.

Results of univariate and multivariate logistic regression.

Variables	Univariate					Multivariate
	β	S.E	Z	P	OR (95%CI)	β	S.E	Z	P	OR (95%CI)
IVF
0(No)					1.00 (Reference)					1.00 (Reference)
1(Yes)	0.88	0.46	1.92	0.054	2.42 (0.98 ~ 5.94)	0.59	0.52	1.13	0.259	1.80 (0.65 ~ 5.03)
Gestational Age	−0.45	0.07	−6.71	< 0.001	0.64 (0.56 ~ 0.73)	−0.45	0.07	−6.60	< 0.001	0.64 (0.56 ~ 0.73)
Mother’s Age	0.05	0.03	1.77	0.076	1.05 (1.00 ~ 1.10)	0.04	0.03	1.44	0.151	1.04 (0.98 ~ 1.11)

OR: Odds Ratio, CI: Confidence Interval

The area under the receiver operating characteristic curve (ROC-AUC) for the nomogram constructed based on the training dataset was 0.80 (95% confidence interval 0.74–0.85) (Fig. 3B), while the ROC-AUC for the validation cohort was 0.80 (95% confidence interval 0.71–0.89). The sensitivity and specificity of the training set and validation set are 85%, 64% and 91%, 55%, respectively (Supplementary Table 1). Furthermore, the calibration curves for both the training and validation sets demonstrated good consistency between the predicted probabilities and actual risks (p > 0.05). The decision curve analysis (DCA) curve is presented in Supplementary Fig. 1, indicating that the use of this nomogram for prediction provides a significant net benefit.

The role of PLT in differentiating severe ROP

The platelet count (PLT) and its related parameters-mean platelet volume (MPV), plateletcrit (PCT), platelet distribution width (PDW) at birth show no significant differences between the ROP group and the non-ROP group. However, at 30 weeks of postmenstrual age, the PLT is significantly lower in the severe ROP group compared to the mild ROP group (p = 0.048), while MPV, PCT, and PDW do not exhibit significant differences (Fig. 4).

Fig. 4.

The changes in platelet-related parameters at different PMA times in mild and severe ROP.

Establishment of a predictive model for severe ROP

Based on treatment received, 110 cases were classified into the mild ROP group and 15 cases into the severe ROP group.After performing Lasso regression on the indicators of the mild and severe retinopathy groups, logistic regression was performed to establish a nomogram model. The results indicated that: gestational age [Odds Ratio (OR): 0.46, 95% Confidence Interval (CI): 0.24–0.87], 5 min Apgar (OR: 0.86, 95% CI: 0.64–1.14), birth weight (OR: 1.00, 95% CI: 0.99–1.00.99.00), and platelet count (OR: 0.99, 95% CI: 0.99–0.99) are significant predictors (results are presented in Supplementary Table 2). This model provides a score to each variable, with the total score derived by summing these individual scores (Fig. 5A). The total score corresponds to the length of a line segment, thereby allowing for the estimation of the probability of developing severe ROP. The chart indicates that factors such as smaller gestational age, lower 5 min Apgar score, lower birth weight, and lower platelet count receive higher scores, thus increasing the risk of severe ROP. The ROC-AUC based on the nomogram was 0.86 (95% confidence interval 0.76–0.96) (Fig. 5B), and the calibration curve demonstrated good consistency between predicted probabilities and actual risks (p > 0.05). The DCA curve is shown in Fig. 5C, indicating that using this nomogram for prediction provides significant net benefits.

Fig. 5.

A: Nomogram model to predict the incidence of severe ROP; B: The ROC-AUC for differentiating mild from severe ROP; C: The calibration curves and DCA curve for differentiating mild from severe ROP.

Discussion

The ROP prediction model developed in this study integrates key parameters including GA, mother’s age, and IVF to predict the risk of ROP early after birth. It achieves risk stratification at a PMA of 30 weeks, with AUC values of 0.80 and 0.86, demonstrating high discrimination ability. This model optimizes the allocation of medical resources by reducing the frequency of invasive screening, while precisely aligning clinical interventions with critical stages of retinal vascular development, such as the vascular occlusion phase and the neovascularization phase, thereby providing a timely window for early intervention.

GA is a critical factor in the risk assessment of ROP, demonstrating a negative correlation with retinal immaturity and susceptibility to oxidative stress¹⁰. Current literature indicates that while maternal age is indeed associated with the incidence of Retinopathy of Prematurity (ROP), a growing body of research suggests that maternal comorbidities, particularly hypertension and diabetes, play a more significant role in the development of ROP^11,12.IVF may directly contribute to ROP risk factors like preterm birth and low birth weight, while high oxygen exposure in the embryonic culture environment may affect retinal vascular development through oxidative stress^13,14. Furthermore, abnormal placental function in IVF pregnancies may exacerbate fetal hypoxia, further promoting the pathological progression of ROP¹⁵. These findings offer new directions for screening high-risk populations.

The negative correlation between platelet (PLT) counts and the severity of retinopathy of prematurity (ROP) contrasts with previous studies that emphasized the association between thrombocytopenia and severe ROP^8,16. This study further reveals that PLT counts in the severe ROP group are significantly lower than those in the mild group at a corrected gestational age of 30 weeks, suggesting that dynamic changes in PLT may serve as a sensitive indicator of disease progression. Platelets release anti-angiogenic factors, such as endostatin, through α-granules to inhibit neovascularization; thus, thrombocytopenia may compromise this protective mechanism^17,18. The significant reduction in PLT counts observed in the severe ROP group at a corrected gestational age of 30 weeks indicates that dynamic monitoring of PLT can effectively reflect the degree of retinal ischemia and hypoxia, providing a basis for determining the optimal timing of treatment.

In comparison to existing research, the AUC of this model (ranging from 0.80 to 0.86) is comparable to that of high-dimensional data models based on machine learning, such as decision trees which have an AUC of 0.83¹⁹. However, it significantly outperforms traditional logistic regression models, which typically exhibit an AUC of less than 0.75²⁰. For instance, the AI-based remote screening model developed by Greenwald et al. achieved an AUC of 0.99²¹; however, it relies on high-resolution fundus images. Additionally, the deep learning system developed by Wu et al., the ROP occurrence prediction model (OC-Net), achieved AUCs of 0.90 and 0.94 in internal and external validation sets, respectively, while the severity prediction model (SE-Net) reached AUCs of 0.87 and 0.88²². Yet, both require the integration of clinical features with retinal images to achieve high-performance predictions. In contrast, this study utilizes only routine clinical data, all of which can be obtained early, thereby enhancing its general applicability.

Based on the literature comparison, the developed prediction model for mild and severe ROP (AUC = 0.86) exhibits several advantages in terms of simplicity, clinical practicality, and performance balance: (1) Minimal parameters: Unlike models that necessitate dynamic monitoring of weight (ROPScore/Children’s Hospital of Philadelphia ROP) or complex medical histories (such as transfusions or mechanical ventilation), this model relies solely on the child’s birth data and the PLT data at 30 weeks of PMA, significantly alleviating the burden of data collection^23,24; (2) Superior performance: The AUC value surpasses that of ROPScore (0.76) and CHOP ROP (0.816) validated in China, while circumventing the issues of low specificity (21.4%) associated with the latter’s pursuit of 100% sensitivity^23,25. In summary, the model developed in this study achieves efficient prediction at a lower clinical cost, providing a more feasible tool for optimizing screening strategies.

The limitations of this study include: (1) The relatively small sample size of the training and validation sets, along with the absence of multi-center external validation, which may affect the model’s generalization ability; (2) The data is sourced from historical medical records, potentially introducing selection bias or information loss, thereby impacting the model’s reliability; (3) In the severe ROP prediction model, the number of events is considerably smaller than the number of variables, which significantly impacts the model’s generalization ability. Future research must prioritize conducting external validation of this nomogram in a larger-scale, multi-center, prospective cohort study.

Conclusion

This study integrates clinical parameters and platelet count as predictive factors to construct a non-invasive tool for the early prediction and risk stratification of retinopathy of prematurity (ROP), thereby providing a novel approach for personalized screening. However, due to the limitations of sample size and the single-center design, future work will focus on validating the generalizability of the model through multi-center prospective cohorts.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

QZ contributed to data curation, methodology, writing of the original draft and editing. JW and DW contributed to investigation, methodology, formal analysis and reviewing. YRL contributed to investigation and data curation. XL and GHL contributed to conceptualization and revising it critically for important intellectual content.

Funding

The research was funded by Science and Technology Development Program of Jinan Municipal Health Commission (No. 2024308009).

Data availability

The data sets generated and analyzed during the current study are not publicly available but are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.