Article Text
PDF
Abstract
Aims To construct and validate an optical coherence tomography (OCT) and optical coherence tomography angiography (OCTA) image model for predicting the occurrence of short-term vitreous haemorrhage (VH) in polypoidal choroidal vasculopathy (PCV) patients.
Methods We retrospectively collected clinical and imaging information from patients diagnosed with PCV at Peking Union Medical College Hospital, Beijing, China, between January 2015 and October 2022. Six different screening strategies, including univariate analysis, multivariate analysis, least absolute shrinkage and selection operator, stepwise logistic regression, random forest and clinical-data-only approach, were used to select variables and build models. The nomogram was constructed based on the model with the best area under the curve (AUC) and was evaluated using receiver operating characteristic curves, calibration curves, decision curve analysis and clinical impact curves.
Results A total of 147 PCV patients were included and randomly divided into a training set (103 patients) and a validation set (44 patients), with an average follow-up time of 17.56±14.99 months. The optimal model that achieved higher AUC in both training and validation sets incorporated seven significant variables identified through univariate analysis: male [OR=2.76, p=0.022], central macular thickness [OR=1.003, p=0.002], the presence of haemorrhagic pigment epithelial detachment (HPED) [OR=6.99, p<0.001], the height of HPED [OR=1.002, p<0.001], the area of HPED [OR=1.16, p<0.001], the presence of multiple PEDs [OR=2.94, p=0.016] and the presence of subretinal haemorrhage [OR=3.11, p=0.011]. A predictive nomogram based on these variables yielded an AUC of 0.896 (95% CI 0.827 to 0.965) in the training set and 0.861 (95% CI 0.749 to 0.973) in the validation set, demonstrating good calibration and clinical usefulness.
Conclusion The proposed OCT/OCTA-based image nomogram, as a novel and non-invasive tool, achieved satisfactory prediction of VH secondary to PCV.
- Hemorrhage
- Retina
- Risk Factors
- Imaging
Data availability statement
Data are available upon reasonable request. All data are available within the manuscript and upon request to the corresponding author.
This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/.
Statistics from Altmetric.com
WHAT IS ALREADY KNOWN ON THIS TOPIC
Vitreous haemorrhage (VH) is one of the most serious and vision-threatening complications of polypoidal choroidal vasculopathy (PCV); thus, an early identification of high-risk individuals is critical. Prognostic indicators have been well reported, but clinical and image risk factors of VH occurrence have been poorly studied.
WHAT THIS STUDY ADDS
A predictive model exhibiting good discrimination, accuracy and clinical usefulness was developed and validated for predicting the occurrence of short-term VH in PCV patients. The model includes seven important variables: sex, central macular thickness, the presence of haemorrhagic pigment epithelial detachment (HPED), the height of HPED, the area of HPED, the presence of multiple PEDs and the presence of subretinal haemorrhage.
HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY
The predictive model provides a direct, non-invasive and robust way for personalised prediction of VH, which could support physicians in patient education and clinical decision-making regarding follow-up frequency or treatment options.
Introduction
Polypoidal choroidal vasculopathy (PCV), one of the pachychoroid spectrum diseases, is characterised by an abnormal branching neovascular network featuring aneurysmal dilations referred to as polyps or polypoidal lesions.1 2 Clinic-based epidemiological studies have reported that the prevalence of PCV ranges from 8% to 62% in cases presumed to be neovascular age-related macular degeneration and is more prevalent in Asian populations.3–6
Vitreous haemorrhage (VH) represents one of the most serious and vision-threatening complications of PCV, occurring in about 15% of cases.7 8 The prognosis remains poor even after pars plana vitrectomy treatment, as choroidal atrophy and scar formation can develop gradually, resulting in irreversible visual impairment.9 10 Hence, the early identification of patients at risk of progressing VH is critical, as it provides clinicians with meaningful information to assess patients’ condition. Additionally, among possible future therapies to prevent PCV from progressing to VH, predicting the recent progression of VH would be valuable in targeting patients who would benefit.
Previous studies have identified several prognostic indicators, including age,11 sex,12 13 systemic diseases,14 15 use of anticoagulants,16 17 history of photodynamic therapy (PDT),12 18 persistent polyps19 and haemorrhagic complications (such as haemorrhagic pigmented epithelial detachment, massive subretinal haemorrhage, haemorrhagic retinal detachment),16 20 which have provided preliminary investigations into the risk factors for VH or submacular haemorrhage secondary to PCV. Despite advancements in understanding the risk factors for VH in PCV, there is a notable absence of a robust predictive model that incorporates multiple risk factors and imaging features to provide a personalised risk assessment for VH in PCV patients. This gap underscores the need for a more sophisticated approach to risk stratification.
Therefore, we intend to use a combination of clinical and ophthalmic image data to develop a model for predicting the occurrence of VH in PCV patients and to assess the predictive accuracy using an independent validation group. With the help of the predictive model, it would be possible for clinicians to identify high-risk individuals and tailor their management strategies accordingly.
Methods
Study design and ethics statement
This retrospective cohort study included patients diagnosed with PCV at Peking Union Medical College Hospital, Beijing, China, between January 2015 and October 2022. Patients were followed consecutively for at least 1 year, receiving comprehensive clinical and imaging examinations during each visit. PCV was diagnosed using the indocyanine green angiography criteria from the EVEREST study.21 22 Only one eye from each patient was included, and the eye with VH occurrence was enrolled preferentially. For patients with bilateral PCV without VH, the eye that had been followed for a longer duration was selected.
Patients were excluded from the study if they met the following criteria: (1) combination of other ocular diseases, such as retinal detachment, retinal artery microaneurysm, central serous chorioretinopathy, retinal vein occlusion, retinal artery occlusion, diabetic retinopathy, uveitis and glaucoma; (2) patients who had undergone intraocular surgery (except cataract surgery); (3) poor quality of OCT/OCTA images (quality index <7); and (4) patients with insufficient medical data or lost to follow-up. The study was approved by the Ethics Committee of Peking Union Medical College and following the Declaration of Helsinki.
Variable collection and image analysis
The data on clinical characteristics were collected through electronic medical records, which were classified as follows: (a) age at diagnosis of PCV, (b) sex, (c) previous treatment (including treatment naïve, anti-VEGF monotherapy, PDT and anti-VEGF combination therapy), (d) systemic history (including hypertension, diabetes mellitus, coronary artery disease, anticoagulant medication use) and (e)axial length.
For image examination, swept source-OCT and swept source-OCTA were performed with one of the following machines: Triton Deep Range Imaging (DRI) OCT device (Topcon Corporation, Tokyo, Japan) or VG200 (SVision Imaging, Ltd., Luoyang, China). OCT protocol included radial scans centred on the lesions and macula (6 or 9 mm, 12 scans). The OCTA protocol consisted of a 6 mm × 6 mm or 9 mm × 9 mm image centred on the lesion and macula for each enrolled eye.
Two fundus ophthalmologists (SYC and XYZ, with 4 and 7 years of experience in fundus imaging, respectively), both blinded to the clinical data, independently reviewed the images and reached a consensus after discussion if disagreements arose. The following OCT and OCTA features were evaluated: (a) central macular thickness (CMT); (b) subfoveal choroidal thickness (SFCT); (c) the presence of multiple pigment epithelial detachment (PED), referring to containing two or more PEDs; (d) the type of the PED, which was classified as serous pigment epithelial detachment (SPED), haemorrhagic pigment epithelial detachment (HPED) and fibrous pigment epithelial detachment (FPED); and (e) the maximum height and area of different type of PED. If there was more than one PED of a particular type, the PED with the largest area would be counted. With OCTA software delineating the region from the RPE to Bruch’s membrane slab, the PED area was calculated by manually demarcating the borders of the PED on en face of OCTA. (d) The presence of sub-retinal pigment epithelium (RPE) ring-like lesion, using the definition according to the Asia-Pacific Ocular Imaging Society PCV Workgroup.23 In brief, it was a round structure with a hyporeflective centre and hyperreflective outline seen under PED (e) the presence of subretinal fluid (SRF), (f) the presence of intraretinal fluid (IRF), and (f) the presence of subretinal haemorrhage (SRH).
Variable selections and model building
We followed a prospectively defined plan to build logistic regression models using various screening strategies, namely, univariate analysis (Model 1), multivariate analysis (Model 2), least absolute shrinkage and selection operator regression (LASSO, Model 3), stepwise logistic regression (Model 4) and random forest (Model 5) to screen candidate variables in the training set. Furthermore, a clinical-data-only approach was used to build Model 6.
Next, binary logistic regression prediction models were constructed based on the different variable selection strategies, and the optimal model was determined based on AUC, accuracy, sensitivity and specificity of both the training and validation cohorts. Then, a nomogram was plotted and the OR and its 95% CI were calculated according to the optimal model.
Performance evaluation of nomogram
The receiver operating characteristic (ROC) curve was drawn to show the discriminative ability of the predictive model, and AUC values were measured to provide a quantitative measure of discriminatory power. The Hosmer–Lemeshow test was used to assess the goodness-of-fit of the model. In contrast to discrimination, calibration curves typically measure the accuracy of predicted probabilities by comparing the agreement between the estimated observed probabilities and the nomogram-estimated probabilities.24 In addition, to determine the clinical usefulness and clinical applicability of the prediction model, the decision curve analysis and clinical impact curve were plotted by quantifying the net benefits at different threshold probabilities in the training set and the validation set.25
Statistical analysis
The continuous variables were analysed by Student’s t-test or Mann–Whitney U-test and were reported as means (SDs) or medians (IQRs). For categorical variables, the X2 test or Fisher’s exact test was performed as appropriate and were reported as frequencies (proportions). All the statistical analyses were done using SPSS (version 26, Chicago, IL, USA) and R software (version 4.3.1, Boston, MA, USA). A significant difference was indicated by p<0.05.
Results
Among 432 patients diagnosed with PCV, 124 were excluded due to a follow-up time of less than 12 months, 103 were excluded because of poor image quality or missing data, 49 were excluded because of comorbidity with other ocular diseases or intraocular surgery history and 9 were excluded because of missing clinical information. A total of 147 patients were ultimately enrolled, including 45 patients developing VH and 102 patients without VH occurrence during the follow-up. The average follow-up duration was 17.56±14.99 months. The final study population was randomly divided into a training cohort (103 cases) and a validation cohort (44 cases) by a ratio of 7:3. Figure 1 depicts the flowchart of the study design.
The flowchart of the study design. HPED, haemorrhagic pigment epithelial detachment; LASSO, least absolute shrinkage and selection operator; PCV, polypoidal choroidal vasculopathy; PED, pigment epithelial detachment.
Patient characteristics
The clinical and OCT/OCTA image features of patients in the training set, validation set and total set are summarised in table 1. There were no significant differences between the VH group and non-VH group in age, systemic history (including hypertension, diabetes mellitus, coronary artery disease and use of anticoagulant medication), axial length, SFCT, the presence of a sub-RPE ring-like lesion, the presence of SPED and its area and height, the area and height of FPED and the presence of SRF and IRF. However, there were significant differences in sex (p=0.004), previous treatment (p＜0.001), CMT(p=0.02), the presence of multiple PEDs(p=0.019), the presence of HPED (p＜0.001), the area of HPED (p＜0.001), the height of HPED (p＜0.001), the presence of FPED (p=0.03) and the presence of SRH (p=0.001).
Demographic characteristics of study participants
Variable selection and predictive model building
Model 1 was constructed using the significance variables (p<0.05) identified from the univariate analysis, including sex, CMT, the presence of HPED, the height of HPED, the area of HPED, the presence of multiple PEDs and the presence of SRH.
The seven significant variables identified from the univariate analysis were subsequently included in a multivariate regression analysis, with those meeting the significance criterion (p<0.05) being finally incorporated into Model 2, including the presence of HPED, the height of HPED, the area of HPED, the presence of multiple PEDs and the presence of SRH.
Model 3 was built using potential factors via LASSO. As shown in online supplemental figure 1, a coefficient plot was generated based on the ln (λ) sequence. With an optimal lambda of 0.068, the LASSO regression analysis identified four non-zero coefficients: CMT, the presence of HPED, the area of HPED and previous treatment.
Supplemental material
Model 4 was constructed with stepwise, forward and backward stepwise regression. Among these, backward stepwise regression had a minimum Akaike’s Information Criterion of 61.7 with nine potential factors, namely, age, previous treatment, CMT, the presence of HPED, the area and height of HPED, the area and height of FPED and the height of SPED.
Model 5 was created with latent variables screened by random forest. The variables were ranked in order of importance (online supplemental figure 2), followed by a tenfold cross-validation method that yielded the model was near peak accuracy when including eleven of the most important variables, which were the area of HPED, CMT, the height of HPED, previous treatment, SFCT, age, axial length, the presence of HPED, the area of FPED, the presence of SRH and the height of FPED.
All clinical data were included to construct Model 6, including age, sex, previous treatment, history of hypertension, history of diabetes mellitus, history of coronary artery disease and anticoagulant medication use.
Next, binary logistic regression models were constructed based on the variables selected by the methods described above, and the ROC curves of the training set as well as the validation set were plotted (online supplemental figure 3). In the training cohort, the AUC of the six models was 0.896 (95% CI 0.827 to 0.965), 0.881 (95% CI 0.815 to 0.947), 0.899 (95% CI 0.820 to 0.978), 0.958 (95% CI 0.903 to 1.000), 0.960 (95% CI 0.897 to 1.000) and 0.760 (95% CI 0.650 to 0.871), respectively, whereas in the validation cohort, the AUC of the six models was 0.861 (95% CI 0.749 to 0.973), 0.831 (95% CI 0.710 to 0.953), 0.811 (95% CI 0.612 to 0.939), 0.799 (95% CI 0.630 to 0.969), 0.811 (95% CI 0.645 to 0.978) and 0.778 (95% CI 0.609 to 0.947), respectively. The detailed performance of the six models, including specificity, sensitivity and accuracy, are shown in table 2.
Predictive efficacy of six different models
The detailed results of univariate analysis and multivariate analysis are presented in table 3. Univariate analysis showed that sex (p=0.022), CMT (p=0.002), the presence of HPED (p＜0.001), the height and area of HPED (p＜0.001 both), the presence of multiple PEDs (p=0.016) and the presence of SRH (p=0.011) were significantly related to the occurrence of VH. At the multivariate analysis, the presence of multiple PEDs (OR 3.87; 95% CI 1.17 to 13.99; p=0.030), the presence of HPED (OR 7.92; 95% CI 1.07 to 57.84; p=0.038), the height of HPED (OR 0.996; 95% CI 0.992 to 1.000; p=0.047), the area of HPED(OR 1.27; 95% CI 1.10 to 1.56; p=0.009) and the presence of SRH(OR 3.34; 95% CI 1.04 to 11.91; p=0.049) were independent risk factors of VH occurrence.
Logistic regression analysis of variables in the training set
Nomogram establishment and evaluation
The model constructed with variables screened by univariate regression analysis was selected as the optimal model based on achieving the highest AUC in the validation set, and a nomogram was developed using this model to predict the occurrence of VH in PCV patients (figure 2A).
Nomogram and validation plots. (A) The nomogram built based on model 1 (variables selected from univariate regression analysis) for predicting the occurrence of vitreous haemorrhage (VH) in polypoidal choroidal vasculopathy patients. (B) Receiver operating characteristic (ROC) curves for the training and validation sets of model 1. The model was constructed using univariate regression analysis to select significant variables, including sex, CMT, the presence of HPED, the height of HPED, the area of HPED, the presence of multiple PEDs and the presence of subretinal hemorrhage. (C) Calibration curves for the training and validation sets of model 1, with P>0.05 in the Hosmer–Lemeshow test, suggesting the nomogram with goodness of fit. (D) Decision curve analysis for the training and validation sets of model 1. The x-axis shows the risk threshold, and the y-axis indicates the net benefit. The red line represents the nomogram for predicting VH. The grey lines indicate that VH is assumed to occur in all patients. The black lines indicate that no VH is assumed to occur in patients. (E) Clinical impact curve for the training and validation sets of model 1. The red curve indicates the number of people who are classified as positive (high risk) by the model at each threshold probability. The blue curve represents the number of true positives at each threshold probability. CMT, central macular thickness; HPED, haemorrhagic pigment epithelial detachment; multi-PEDs, multiple pigment epithelial detachment (with two or more PEDs present); SRH, subretinal haemorrhage. AUC, area under curves
Satisfactory predictive performances of the nomogram were obtained with an AUC of 0.896 in the training set and an AUC of 0.861 in the validation set (figure 2B). Furthermore, favourable calibration curves applied to the training and validation cohorts were confirmed for the proposed nomogram (figure 2C), and the p value was 0.543 for Hosmer–Lemeshow test, both indicating optimal goodness of fit. In addition, the decision curve analysis showed that neither in the training group nor in the validation group, the nomogram in predicting the occurrence of VH in PCV patients could yield clinical net benefits (figure 2D). In addition, clinical impact curves show that the overall net benefit of the nomogram was higher and had a greater impact on patient outcomes within the wide and practical range of threshold probabilities (figure 2E).
Discussion
The progression of VH in PCV patients is indicative of severe visual impairment and a poor prognosis. Although many studies have investigated risk factors for VH in PCV patients, these studies often suffer from small sample sizes, produce inconsistent results and fail to develop effective predictive models. In this retrospective cohort study, we acquired imaging features through OCT and OCTA and screened the variables using multiple strategies to construct an optimal model with the best predictive efficacy in evaluating the probabilities of VH secondary to PCV. Furthermore, we developed and validated a nomogram based on the optimal model, which incorporated sex, CMT, the presence of HPED, the height and area of HPED, the presence of multiple PEDs and the presence of SRH. Notably, the nomogram provides a direct, non-invasive and robust way for personalised prediction of VH with high accuracy, calibration and clinical application.
Numerous studies have attempted to identify the underlying characteristic risk factors of VH in PCV patients. The effects of anticoagulant medications on VH stand to be controversial. Retrospective studies have shown that daily antiplatelet or anticoagulant medication usage was significantly associated with an increased risk of SRH or VH in patients with age-related macular degeneration.16 17 Conversely, other studies found no notable correlation between anticoagulants and haemorrhagic complications among PCV patients.12 26 27 Our univariate analysis revealed that anticoagulants were not significantly associated with VH (p=0.986). This discrepancy could be attributed to differences in study designs and patient populations, the former studies included older age-related macular degeneration in patients with average ages of 70.816 and 83.1 years,17 whereas the latter focused on younger PCV patients with an average age of 62–69 years.12 26 27 Additionally, variations in baseline health status and comorbidities of the study populations may also contribute to these differing results. In addition, Rishi et al12 reported that women were associated with a reduced risk of haemorrhage, which is consistent with our conclusion that men are at higher risk of VH (p=0.022), but the connection between sex and VH could not be elucidated.
Studies have also been devoted to applying images to predict disease prognosis. Wu et al7 and Shin et al16 reported that a larger size of submacular haemorrhage was significantly positively related to a greater risk of VH. Others found that thicker SRH or subretinal elevation detected via ultrasonography was associated with poor visual prognosis,9 10 28 which was compatible with our findings. Univariate analysis revealed that the presence of SRH was significantly associated with the possibility of VH occurrence (OR= 3.11; p=0.011). This is consistent with clinical observations and experience, indicating that the high rate of haemorrhagic complications observed in PCV patients may be attributed to the presence of branching vascular network with aneurysmal polypoidal dilations, which are prone to rupture and haemorrhage.
Additionally, our study found that multiple PEDs and HPED were critical indicators of VH, echoing findings from previous reports. A retrospective case-control study of 98 eyes (36 eyes with breakthrough VH and 62 eyes without VH) showed significantly higher percentages PED, HPED, haemorrhagic retinal detachment and haemorrhagic choroidal detachment in the VH group compared with the non-VH group.20 Another retrospective case-control study involving 722 patients (103 eyes with breakthrough VH and 619 eyes without VH) found HPED present in 81% of patients in the VH group, significantly higher than the 17% in the non-VH group (p＜0.001).8 The role of PED as an important biomarker in chorioretinal disease has garnered widespread attention,29 and advances in imaging technology have facilitated quantitative assessments of PED in evaluating various diseases.30 31 In this study, HPED was identified as a risk factor for VH, and incorporating measurements of the height and area of HPED in the nomogram could enhance the accuracy of the model.
VH is a troublesome complication of PCV, making the identification of at-risk individuals a crucial and urgent task. To date, only one prediction model has identified risk populations for developing VH based on systemic factors in PCV patients. Liu et al26 developed a model based on systemic parameters, including activated partial thromboplastin time, aspartate aminotransferase/alanine aminotransferase ratio and white blood cell count, which reported an AUC of 0.723, sensitivity of 0.609 and specificity of 0.773. However, this study lacks validation with an independent dataset, potentially leading to overfitting in the analysis32 and poor generalisability of the model. In our study, the use of clinical and imaging-based variable model contributed to an AUC over 0.85 in both the training and validation cohorts. Additionally, we constructed a clinical-data-only model (Model 6), which achieve an AUC of 0.760 in the training set and 0.778 in the validation set. These underscore the challenges and limitations of predictive models based solely on clinical parameters, which may not fully capture the nuanced characteristics of VH in PCV patients as effectively as imaging-based models do.
In addition to conducting validation to prevent overfitting, a strength of our study is its pioneering role in developing the first imaging-based model of the VH occurrence in PCV patients, demonstrating satisfactory predictive performance in terms of discrimination, accuracy, calibration and clinical application. A nomogram was developed to guide ophthalmologists in clinical decision-making concerning follow-up frequency and treatment options. However, there were still some limitations. First, it was a single-centre study that lacked external validity, and the results might not be generalisable for the Western population. Second, due to the difficulty of data collection, as massive VH can affect the quality of image reports, the small sample size of our study posed a challenge in performing subgroup analyses considering the timing of the VH occurrence. Finally, OCT and OCTA images were collected using different scanners with varying scanning parameters, which may result in heterogeneity bias. Therefore, a prospective multicentred study with larger samples is needed to improve the model in the future.
Conclusion
In conclusion, this study demonstrated that non-invasive imaging parameters obtained from OCT and OCTA serve as powerful predictors of whether VH occurs in PCV patients. The nomogram integrates and visualises these predictive factors, aiding physicians in clinical decision-making and patient education.
Data availability statement
Data are available upon reasonable request. All data are available within the manuscript and upon request to the corresponding author.
Ethics statements
Patient consent for publication
Ethics approval
Not applicable.
References
Supplementary materials
Supplementary Data
This web only file has been produced by the BMJ Publishing Group from an electronic file supplied by the author(s) and has not been edited for content.
Footnotes
Contributors SC designed the study, collected data, performed the statistical analysis and drafted the manuscript. XZ collected and analyzed the data. QZ and LM participated in revising the manuscript. YC (guarantor) accepts full responsibility for the finished work and/or the conduct of the study, had access to the data and controlled the decision to publish.
Funding This work was supported by the National Natural Science Foundation of China under Grant 82271112; Beijing Natural Science Foundation Beijing-Tianjin-Hebei Basic Research Funds under Grant J200007.
Competing interests None declared.
Provenance and peer review Not commissioned; externally peer reviewed.
Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.