Article Text
PDF
Abstract
Introduction Diabetic retinopathy (DR) is a leading cause of preventable blindness among working-age adults, primarily driven by ocular microvascular complications from chronic hyperglycemia. Comprehending the complex relationship between microvascular changes in the eye and disease progression poses challenges, traditional methods assuming linear or logistical relationships may not adequately capture the intricate interactions between these changes and disease advances. Hence, the aim of this study was to evaluate the microvascular involvement of diabetes mellitus (DM) and non-proliferative DR with the implementation of non-parametric machine learning methods.
Research design and methods We conducted a retrospective cohort study that included optical coherence tomography angiography (OCTA) images collected from a healthy group (196 eyes), a DM no DR group (120 eyes), a mild DR group (71 eyes), and a moderate DR group (66 eyes). We implemented a non-parametric machine learning method for four classification tasks that used parameters extracted from the OCTA images as predictors: DM no DR versus healthy, mild DR versus DM no DR, moderate DR versus mild DR, and any DR versus no DR. SHapley Additive exPlanations values were used to determine the importance of these parameters in the classification.
Results We found large choriocapillaris flow deficits were the most important for healthy versus DM no DR, and became less important in eyes with mild or moderate DR. The superficial microvasculature was important for the healthy versus DM no DR and mild DR versus moderate DR tasks, but not for the DM no DR versus mild DR task—the stage when deep microvasculature plays an important role. Foveal avascular zone metric was in general less affected, but its involvement increased with worsening DR.
Conclusions The findings from this study provide valuable insights into the microvascular involvement of DM and DR, facilitating the development of early detection methods and intervention strategies.
- diabetic retinopathy
Data availability statement
No data are available.
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
The relationship between microvascular changes in the eye and disease progression poses complex challenges that cannot be interpreted fully using classic statistical methods.
WHAT THIS STUDY ADDS
Our data indicates that the ocular microvascular involvement of different stages of diabetic retinopathy is complex and not fully understood. When discriminating between diabetes mellitus without diabetic retinopathy and healthy eyes, microvascular parameters extracted from the choriocapillaris were the most important. This indicates that the choroid is affected very early in diabetes mellitus. As the disease progresses, the deep retinal vascular plexus becomes the most important factor in the discrimination of diabetic retinopathy and diabetes mellitus without diabetic retinopathy. In the later stages of disease, the differences in the superficial retinal vascular plexus are the most pronounced.
HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY
The findings from this study provide valuable insights into the microvascular involvement of diabetes mellitus and diabetic retinopathy, facilitating the development of early detection methods and intervention strategies.
Introduction
Ocular microvascular complications resulting from chronic hyperglycemia contribute to the development and progression of diabetic retinopathy (DR), a leading cause of preventable blindness among working-age adults.1 While conventional funduscopic imaging is commonly used to observe DR-related lesions such as microaneurysms, hemorrhages, neovascularization, and hard exudate deposits, it has limitations in identifying and quantifying microvascular changes that occur before clinically evident DR.2 Conventional imaging modalities are typically limited in their ability to visualize microvasculature with the required detail, resolution, and depth information3 to study the microvascular changes accompanying the development of diabetes mellitus (DM) and non-proliferative DR (NPDR). However, with the impending escalation of the burden of DM and DR, there is a pressing need to deepen our understanding of how their microvascular involvement relates to ocular complications.
Optical coherence tomography angiography (OCTA) holds promise in providing valuable insights as it enables the visualization of the ocular vascular beds with details. Previous studies have used OCTA imaging to detect retinal microvascular abnormalities associated with DR progression, such as increased vessel tortuosity, enlarged and non-circular foveal avascular zones (FAZs), and capillary dropout.4 Furthermore, swept-source optical coherence tomography (OCT) with longer wavelength has been used to better identify choroidal microvascular changes related to DR, including choroidal infarcts and impaired flow.5
Multivariate linear regression models,6 using a combination of retinal and choroidal microvascular parameters, aid the discrimination of different DR severities. However, linear regression assumes a linear, independent relationship between these parameters which may not hold true due to the complex relationship between disease progression and microvascular features. Adopting a non-parametric machine learning (ML) approach, without relying on the assumption of linearity, is better suited for probing this non-linear relationship, as well as to account for complex interactions between parameters that are absent in linear models. Several studies have demonstrated the efficacy of employing ML7 and deep learning8 9 techniques for the automated diagnosis of DR. A significant amount of research has gone into feature engineering for these techniques. Sandhu et al demonstrated that combining parameters extracted from both OCTA and OCT images improves the discriminatory power of random forest (RF) classification for the automated diagnosis of NPDR.10 Carrera-Escalé et al demonstrated that radiomics extracted from OCT and OCTA images can be used as predictors in logistic regression, linear discriminant analysis, support vector classifier (SVC)-linear, SVC-radial basis function, and RF models to identify patients with DM, DR, and referable DR.11 However, there has been limited research into the decision process of the models. Interpreting ML models is challenging given their complexity. SHapley Additive exPlanations (SHAP),12 a metric relying on shapley values, may provide local interpretability to understand the overall importance of individual features (eg, retinal perfusion) in DR prediction, as well as how the interactions of features contribute to the DR prediction of the eye.
This study aimed to identify the importance of microvascular parameters associated with DM and different DR severities to facilitate the development of early detection methods and intervention strategies. To accomplish this, we employed an ML approach using a RF classification to discriminate among four different disease classes: healthy, DM no DR, mild DR, and moderate DR. SHAP values were used to determine the importance of the parameters.
Research design and methods
Study participants
Participants from two separate studies, SIENA (Singapore Imaging Eye Network) and DYNAMO (Diabetes Study in Nephropathy and other Microvascular Complications), were included in this work. The studies took place from April 2018 to July 2019, at the Singapore National Eye Centre. All participants provided written informed consent before their inclusion in the study.
All participants in the DM groups had type 2 DM. The severity of DR was assessed using two-field fundus photography and the Early Treatment Diabetic Retinopathy Study DR grading scale.13 The study included participants with varying degrees of DR (DM no DR, mild DR, and moderate DR); participants with severe DR were excluded. Additionally, we excluded eyes with diabetic macular edema due to the potential for associated morphologic changes to introduce artifacts in the imaging of deeper layers.14 The inclusion criteria for all participants were as follows: age 40 years or older, absence of other ocular pathologies such as glaucoma, age-related macular degeneration, significant media opacity, and proliferative DR (PDR).
This study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology reporting guidelines to ensure transparent and comprehensive reporting of the study findings.
Optical coherence tomography angiography
Trained ophthalmic technicians performed scans on all participants using an SS-OCTA system (PlexElite 9000, Zeiss Meditec, Dublin, California, USA). This system uses a wavelength scanning laser with a light source at a wavelength of 1050 nm. The scanning speed depends on the swept source’s scanning rate, which is 100 000 A-scans/s. The system provides high-resolution imaging with axial and lateral resolutions of 6.3 µm and 20 µm in tissue, respectively.
A standardized scanning protocol of 3×3 mm2 centered at the fovea was applied for data acquisition. Each data volume consisted of 300 A-scans and 300 B-scans. Each B-scan was repeated four times using an optical microangiography algorithm to generate OCTA images. An integrated line scanning ophthalmoscope eye tracker minimized motion-related artifacts during data acquisition. Automated segmentation of retinal layers and retinal pigment epithelium (RPE) was performed using review software (Zeiss Meditec). Manual corrections were applied to address any inaccurate automatic segmentation. The choriocapillaris layers were segmented with a depth range between 31 µm and 39 µm beneath the RPE.
Scans that met any of the following criteria were excluded from further analysis: poor image clarity, weak local signals due to obstacles such as vitreous floaters, and excessive motion artifacts.
Retinal and choroidal parameters
The FAZs in the superficial capillary plexus (SCP) and deep capillary plexus (DCP) were automatically segmented using U-Net, a convolution-based deep learning network.15 Additional information relating to the FAZ segmentation model can be found in the online supplemental tables 1 and 2. The perimeter and area of the SCP FAZ were calculated from the masks, while the DCP FAZs were omitted due to the less distinct boundaries and dense vessel patterns in the DCP, unlike the well-defined radial patterns observed in the SCP.16
Supplemental material
Annulus masks with a width of 500 µm were generated around the SCP and DCP FAZs. Subsequently, the retinal angiograms were binarized using a global threshold, assigning a value of 1 to the perfused area and 0 to the background. The perfusion density (PD) was calculated as the perfused area per total annulus area. Consequently, the binarized perfusion map was skeletonized to reduce vessel diameter to a single pixel, enabling the calculation of vessel density (VD) as the ratio of vessel length to the total annulus area.6
For the choriocapillaris layer, the flow deficit density (FDD) of small and large flow deficits was calculated. Flow deficits represent areas within the choriocapillaris that exhibit a low OCTA signal, indicating hypoperfusion.16 These deficits, with sizes greater than the intercapillary distance, were used to calculate the FDD by measuring the ratio of the deficit area to the total area of the image. The choriocapillaris angiograms were extracted and compensated using corresponding morphological images, followed by binarization using a threshold of mean-SD.6 The size of each flow deficit was then calculated, resulting in two FDD values: one including all sizes and another with a size-selective threshold of 800 µm2. A figure illustrating the feature extraction methodology can be found in the online supplemental figure 1.
Statistical analysis
Multicollinearity between the parameters was measured using the variance inflation factor, and a threshold of ≤5 was considered. It was found that there was high multicollinearity between the SCP VD and PD and the SCP FAZ area and perimeter, making feature selection necessary. A final feature combination of the SCP FAZ perimeter, SCP VD, DCP VD, large flow deficits, and small flow deficits was selected as they delivered the highest classification performance.
The parameters were used as predictors for four binary classification tasks related to the identification of DM and DR severity. Each task was crafted to answer different questions relating to the microvascular involvement of DM and DR in the eye:
Comparison 1: DM no DR versus healthy.
Comparison 2: mild DR versus DM no DR.
Comparison 3: moderate DR versus mild DR.
Comparison 4: any DR versus no DR.
All parameters were pre-adjusted to account for changes in the microvasculature that occur with age. Each parameter was fit to a linear regression model with age, the original values of the parameters were then replaced with the difference between the residuals and original values.
A stratified fivefold cross-validation technique was used for the training and testing of the classification models in order to account for the class imbalance and ensure a fair trade between the models’ bias and variance. The hyperparameters of the models were fine-tuned using a grid-search approach. Details relating to the fine-tuning can be found in the online supplemental table 3.
An RF classification algorithm was implemented using the scikit-learn Python library17 and was employed for all classification tasks. It is a supervised ML technique that forms an ensemble of decision trees, with each tree trained independently using a random selection of a subset of the data.18 This particular technique was chosen for two reasons: its capability to effectively capture intricate relationships within the input data and its ability to tackle challenges associated with limited sample sizes without the risk of overfitting.19 Traditionally, feature importance in RF models is studied using the Gini index, which calculates which features provide the greatest contribution across the entire population.20 However, previous studies have shown that this method is vulnerable to multiple biases making it unreliable in many scenarios.21 Consequently, we calculated the feature importance of the parameters using SHAP.12 22 SHAP is a local feature importance method; a value for each of the predictors is calculated for each prediction. These values are representative of the contribution of the predictor to the final prediction. As the performance of the model does not impact the SHAP values, the values from each of the testing folds were combined and reported in the SHAP figures. Furthermore, partial dependence plots (PDPs) with SHAP values were employed to analyze interactions between the parameters.
Results
Patient characteristics
The study included a healthy group comprising 196 eyes from 119 participants. Participants with DM were categorized into three groups based on DR severity: no DR (120 eyes from 79 participants), mild DR (71 eyes from 53 participants), and moderate DR (66 eyes from 44 participants). The patient characteristics are summarized in table 1. There was no significant difference in the diastolic blood pressure (BP) (p=0.982). However, there was significant difference in age (p<0.001) in the DM no DR versus healthy, moderate DR versus DM no DR, and moderate DR versus mild DR groups, in body mass index (p<0.001) between the healthy versus DM no DR, and DM no DR versus moderate DR groups, and in the hypertension (p<0.001) and systolic BP (p<0.001) in the DM no DR versus healthy group. Blood test and duration of DM data were collected only in the DYNAMO Study; this information can be found in the online supplemental table 4.
Study participant characteristics by diabetes mellitus (DM) and diabetic retinopathy (DR)
Microvascular parameters
Figure 1 presents a flow chart of the study. Representative color fundus images (figure 1A) and OCTA images from the retina (figure 1B) and choriocapillaris (figure 1C) from healthy, DM no DR, mild DR, and moderate DR eyes are shown. The OCTA scans in the macular region provide a detailed visualization of the microvasculature. The choriocapillaris and retinal vessels were significantly impacted by increasing DM or DR severities (p<0.001). A detailed quantitative summary is provided in table 2.
Study protocol: Representative images from diabetic retinopathy using (A) color fundus photography and individual OCTA plexuses, (B) superficial capillary plexus, (C) deep capillary plexus, and (D) choriocapillaris plexus. Microvascular features were extracted from the OCTA images and used as predictors for a machine learning classifier to discriminate between healthy, DM no DR, mild DR, and moderate DR eyes. The feature importance of the microvascular parameters was calculated from the classifiers built. DM, diabetes mellitus; DR, diabetic retinopathy; OCTA, optical coherence tomography angiography.
Quantitative metrics from retinal perfusion, superficial FAZ, and choriocapillaris
Expectedly, all the OCTA metrics were significantly associated with DM or DR stages (p≤0.002), and boxplots (online supplemental figure 3) show distribution of the parameters at different severities. The Pearson’s R correlation between each parameter is summarized in a correlation matrix (online supplemental figure 4). There was a strong correlation (r>0.75) between the large flow deficits and small flow deficits and a moderate correlation (0.5<r<0.75) between the SCP and DCP VD. All other parameters were only weakly correlated (r<0.5). Specifically, in the choriocapillaris, an increase in the density of large flow deficits was seen in comparison 1 (p<0.001) and comparison 2 (p<0.001) but not in comparison 3. An increase in the density of the small flow deficits was found in comparison 1 (p<0.001). SCP and DCP VD declined with the increasing severity of DR and was significantly different in comparison 2 (p<0.001) as well as comparison 3 (p<0.001). In comparison 1, the SCP VD was reduced (p<0.001), but not the DCP VD. An increase in the FAZ perimeter was observed in comparison 1 (p<0.001) as well as DM no DR eyes and moderate DR eyes (p<0.001).
ML and feature importance
The performance metrics of the ML model including F1 score, precision, and recall are reported in table 3. The sequential severity comparison yielded an F1 score between 0.6 and 0.7, while the comparison between no DR versus any DR yielded a high F1 score of 0.840±0.024.
Precision, recall, and F1 score mean and SD values for DR severity classification
The feature importance was calculated using SHAP at each testing fold, and the SHAP plots as well as the ranking trend for the parameters are shown in figure 2. The SHAP values revealed the varying involvement of the microvascular parameters at different stages of DM and DR. The large flow deficit parameter was the most important for the differentiation in comparison 1, its importance progressively decreased with the increase in disease severity. The small flow deficit parameter was overall less important than the large flow deficits, but still more important in comparisons 1 and 2 than in comparison 3. The SCP VD was consistently highly ranked. It was determined to be the most important parameter in comparison 3, and second only to the large flow deficits in comparison 1. Interestingly, it was the least important parameter in comparison 2, where the DCP VD dominated the transition. The FAZ perimeter was consistently ranked low in the early phases of DM and DR, but its importance increased to second in comparison 3.
Feature importance results: (A) SHAP value plots for the four classification tasks: DM no DR versus healthy, mild DR versus DM no DR, moderate DR versus mild DR, and any DR versus no DR. Each of the points on the SHAP plots is representative of the calculated SHAP value of the parameter for a single testing sample, the color denotes the value of the parameter. Negative SHAP values show that the parameter is pushing the predicted probability towards a class 0 prediction and vice versa. The y-axis arrangement of the parameters is arranged based on the absolute sum of the SHAP values. The change in the ranking of the parameters demonstrates that different microvascular parameters are important for the classification tasks. (B) Ranking of the parameters based on SHAP values: The large and small flow deficit parameters were highly ranked when detecting the presence of DM and mild DR, the deep vessel density was the most important in distinguishing DM from mild DR, the superficial vessel density was consistently highly ranked except in the mild DR versus DM no DR classification and the FAZ perimeter was consistently ranked low except in the moderate DR versus mild DR classification. DM, diabetes mellitus; DR, diabetic retinopathy; FAZ, foveal avascular zone; SHAP, SHapley Additive exPlanations.
The PDPs with SHAP values can be found in the online supplemental figure 2. In comparisons 1 and 4, non-linear interactions were observed to be impacting all of the parameters. In comparison 2, it is observed that the only parameter that had no definite relevant interaction was the large flow deficits parameter. However, for comparison 3, it was difficult to gauge whether interactions were impacting the SHAP values because of the limited sample size.
Discussion
In this study, we evaluated the importance of retinal and choroidal microvascular parameters in different stages of DM and DR, with SHAP values derived from an RF model. Retinal and choroidal microvasculature are significantly affected by DM and DR severities, and their involvements in different stages vary. Specifically, in the choriocapillaris, an increase in the density of large flow deficits was the strongest parameter discriminating healthy eyes from eyes with DM no DR. In eyes with DM, VD in the SCP and DCP became the most important discriminative factors with increasing severity of DR. Overall, these findings provide insights into the relationship between microvascular parameters, DM, and DR severity, and suggest the potential utility of ML techniques in classifying and understanding the progression of these conditions.
We found that alterations in the choriocapillaris dominated the difference between healthy and DM no DR eyes. Choriocapillaris impairment causes small flow deficits to merge into large flow deficits thereby reducing the number of flow deficits but increasing their size.23 Using swept-source OCTA, choriocapillaris flow impairment quantified by the density and size of flow deficits was previously reported in DM no DR eyes.6 24 Histological studies on the human eye have demonstrated both focal and diffuse choriocapillaris dropout in diabetic eyes without DR.25 Inflammation plays a crucial role in choriocapillaris dropout in which activated leukocytes bind to increased expression of adhesion molecules on the endothelium under hyperglycemia environment, leading to leukostasis and subsequent occlusion of capillary lumen.26 The expression of leukocyte adhesion molecules was reported low in retinal vessels compared with choroidal vessels, which may support the hypothesis of an earlier involvement of choriocapillaris in DM no DR eyes.27 Additionally, the segmental vascular pattern of choriocapillaris, with no anastomosis to its neighboring area, makes it more susceptible to ischemia.28 A genetic mouse model of diabetes (Ins2Akita) showed that choroidal blood flow deficits, as measured by MRI, precede alterations of retinal perfusion and visual function.29 Moreover, dysfunction of outer retina, including RPE and photoreceptor, is gaining more attention in early DR,30 which abuts and is largely nourished by the choriocapillaris. This may reinforce our finding that although retinal neural damage occurs in DR development and progression, choroidal vasculopathy may precede neuropathy and retinal vascular changes representing the early sign on DM-related microvascular changes in the eye.
The DCP VD was found to be an essential feature in the differentiation of mild DR eyes from DM no DR eyes. This verifies previous findings that the deep retinal vessels are more likely to be impacted in eyes with mild DR compared with no DR.31 Similarly, another study found that capillary flow deficit was dominant in the DCP rather than the SCP in mild DR eyes compared with no DR eyes.32 Hence, this finding is compatible with previous hemodynamic studies. The retinal oxygen metabolism is altered in DM and oxygen extraction was found to decrease with the worsening of DR severity.33 The resulting hypoxia may initiate retinal dilatation of the SCP to meet the ganglion cell layer’s metabolic demands.34 This hypothesis is supported by the observation that total retinal blood flow as measured using laser Doppler velocimetry combined with vessel diameter measurement increases in early DM.35 This results in the shunting of blood through the SCP rather than traversing the intermediate and deep capillary plexus, which was also demonstrated in diabetic rats.36 Consequently, this secondarily decreased flow in DCP microvasculature may fall below the threshold for detection by OCTA, manifesting as reduced DCP VD in mild DR eyes.
The SCP VD is the most differentiable parameter in separating eyes with any DR from no DR and distinguishing eyes with moderate DR from eyes with mild DR. This means that a distinct change in the SCP VD parameter is observed in the eyes with DR signs, especially in moderate DR eyes. Ashraf et al found a significant difference in SCP metrics between mild and moderate NPDR eyes but not between eyes with no DR and mild DR.31 Two studies found that the adjusted flow index, an indirect measure of flow velocity, in the SCP was significantly higher in the DM no DR group but lower in the DR group than in the healthy eyes.37 This initial preservation of blood flow in SCP could be due to the dilation of superficial capillary vessels in response to autoregulation.38 With the worsening of diabetes, an increase in blood viscosity and a decrease in vessel wall elasticity over-ride the earlier compensatory dilatation,39 leading to a decline in SCP flow with DR severity. The reduction in SCP VD could also result from the loss of retinal ganglion cells (RGCs) and their axons.40 This inner retinal neurodegeneration in diabetes has been shown by several groups.41 42 Reduced RGCs would lead to reduced metabolic demand, disrupting neurovascular autoregulation and leading to a consequent reduction in SCP blood flow.42
Alteration in the FAZ perimeter was more distinctly seen in the moderate DR stage but not in the earlier phases due to large individual variations. Enlargement of the FAZ in DR is related to micro-infarction within the surrounding vascular arcades.43 Thus, SCP dropout is paralleled by FAZ widening. Conrath et al also found that using high-quality fundus fluorescein angiography, the FAZ area was significantly greater in pre-PDR than in background DR.44 Likewise, greater irregularity in the FAZ was observed in the moderate-severe NPDR group relative to controls but not with mild DR using OCTA.45 Another study by Lu et al also reported a significantly larger mean FAZ area in the mild-to-moderate NPDR group than in the DM no DR group.46 However, these studies combined moderate DR eyes either with mild or severe DR and were not separately assessed.
Interactions between the parameters were observed in the PDPs (online supplemental figure 2). Each testing sample is represented by a point on the plot. The horizontal axis is the parameter being impacted, the left vertical axis is the SHAP value for the parameter, and the right vertical axis, which is shown on the plot by color, is the parameter found to be most likely interacting with the parameter being studied. If there is no relevant interaction between the parameters, the coloring of the points on the plot will be uniformly distributed. If there is an interaction between parameters, shades of a different color on the plots will be intersecting. An example of this is the FAZ perimeter in the any DR versus no DR classification task (online supplemental figure 2A) which was determined to be most likely interacting with the SCP VD parameter. There is a definite pattern of intersection in the plot which demonstrates an interaction that is relevant. The plot shows that when the SCP VD is low, the contribution of the FAZ perimeter to the prediction (absolute SHAP value of FAZ perimeter) is higher. When the SCP VD is high, the contribution of the FAZ perimeter is lower. Hence, the SCP VD is changing how the FAZ perimeter contributes to the predicted probability. On the other hand, an example of a parameter that is not interacting with other parameters is the large flow deficits in the mild DR versus DM no DR classification task (online supplemental figure 2B). The parameter that was determined to be most likely interacting with it was the DCP VD; however, the distribution of the color of the points on the graph was relatively uniform, in other words, there was no definite pattern of intersection in the color of the points on the plot.
Study limitations
Our study does have a few notable limitations. It is a cross-sectional study, and a longitudinal study would be helpful in understanding the microvascular importance in disease progression. Eyes with advanced DR, such as severe NPDR and PDR, were not included as we had access to a limited cohort size and a majority of the patients with PDR had received treatment. The duration of DM and blood test data were collected for the DYNAMO dataset but not SIENA and could not be included. In future studies, additional clinical data will be collected. The FAZ perimeter was not scaled to the axial length of the eye. The intermediate capillary plexus was not evaluated separately, which could potentially be involved in DR progression. We used a limited field of the macular and hence, the peripheral retinal changes were not assessed.47 There was a significant difference in age between the healthy and DM no DR eyes. Although metrics’ age dependency was pre-adjusted, its interaction with other OCTA parameters may contribute to the prediction. Differences in BP may be attributed to the inclusion criteria of the present study. Isolated systolic hypertension, which is an elevation in systolic but not diastolic BP, is the most prevalent type of hypertension in people 50 years old and over.48 There is a close relationship between DM and cardiovascular disease, with hypertension being very prevalent in patients with type 2 DM.49 Systolic BP in particular has been shown to be associated with DM complications.50
Conclusion
Feature importance derived from an RF prediction model effectively detected the retinal and choroidal microvascular involvement in eyes at different DM and DR stages. It gave a comprehensive insight that early microvascular change may occur in the choroid and shift toward retina in later stages of DR. Hence, therapeutic targets for different stages of DR should be adaptive. In the future, the current framework can be extended to other features from multiple imaging modalities and studying the combined feature importance from both imaging and biomarker parameters relating to the occurrence and development of DR.
Data availability statement
No data are available.
Ethics statements
Patient consent for publication
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
Presented at This work was presented as a poster presentation at the International Conference on AI in Medicine, Singapore, August 5–7, 2023.
Contributors TSA and BT were involved in the conception, design, and conduct of the study and the analysis and interpretation of the results. TSA wrote the first draft of the manuscript, and all authors edited, reviewed, and approved the final version of the manuscript. TSA is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Funding This work was funded by grants from the National Medical Research Council (CG/C010A/2017_SERI; OFLCG/004c/2018-00; MOH-000249-00; MOH-000647-00; MOH-001001-00; MOH-001015-00; MOH-000500-00; MOH-000707-00; MOH-001072-06; MOH-001286-00), National Research Foundation Singapore (NRF2019-THE002-0006 and NRF-CRP24-2020-0001), A*STAR (A20H4b0141), the Singapore Eye Research Institute & Nanyang Technological University (SERI-NTU Advanced Ocular Engineering (STANCE) Program), and the SERI-Lee Foundation (LF1019-1) Singapore.
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.