Retinal imaging and supervised learning predict hospitalizations and kidney and heart-lung damage in sickle cell disease

Key Points

• •Machine learning on retinal imaging and basic laboratory tests predicted future hospitalizations and organ damage in SCD.
• •Retinal imaging is a noninvasive, promising predictor of morbidity in SCD.

Visual Abstract

Abstract

Sickle cell disease (SCD) is a single-gene illness that causes painful vaso-occlusion, debilitating organ damage, and early mortality. Its clinical course is variable, but current prognostic tools have large data requirements. Understanding prognosis is important for clinicians and patients in selecting therapies, particularly curative therapies. Retinal imaging may provide a noninvasive indicator of vaso-occlusion and risk of organ damage in SCD. This prospective observational cohort study included 150 individuals (aged >15 years) living with SCD. Retinal optical coherence tomography angiography and basic laboratory tests were performed at 6-month intervals over a median of 450 days (interquartile range, 295-744) of follow-up. Primary outcomes were hospitalization within 12 months, kidney damage (urine albumin-to-creatinine ratio of >100 and increased by ≥15 at next study visit), and heart-lung damage (N-terminal pro–brain natriuretic peptide level of >160 pg/mL and increased by ≥15 pg/mL at next study visit). We processed retinal image metrics for perfusion and vascularity using established methods. We extracted retinal image features using a Swin transformer and created ensemble models to predict outcomes. Random forest models predicted hospitalization within 12 months with an area under the receiver operating characteristic (AUROC) curve of 0.717 (standard deviation [SD], 0.053; five-fold cross-validation). We predicted future kidney damage with an AUROC of 0.881 (SD, 0.083) and heart-lung damage with an AUROC of 0.866 (SD, 0.106). Ensemble models, including transformer-derived image features, processed retinal image metrics, and laboratory tests, performed best. Using supervised machine learning on noninvasive retinal imaging and basic laboratory tests, we predicted future SCD hospitalizations and organ damage. This method shows promise for prognostication in SCD.

Introduction

Sickle cell disease (SCD) is the most common hereditary hemoglobinopathy worldwide, affecting 515 000 newborns annually and with a global prevalence of 7.74 million individuals.¹ It is caused by a mutation in the β-globin gene, leading to conformational changes in hemoglobin (Hb) during deoxygenation, which causes red blood cells to “sickle” and occlude blood vessels. In many individuals, this leads to lifelong multisystem complications, including organ damage and failure, impaired quality of life, and early mortality, but outcomes are highly variable. Predicting need for hospitalization and organ damage is of utmost clinical importance in SCD because it can help guide therapeutic and clinical decisions; however, prediction has thus far been challenging in multiple cohorts.2, 3, 4 In particular, heart-lung and kidney damage are both strongly associated with mortality in SCD,5, 6, 7 and the need for hospitalization is an important patient-centered outcome.

Therapeutic choices in SCD each have unique risk-benefit profiles. From the first description of SCD in 1910, there were no approved therapies for treatment until hydroxyurea was approved by the US Food and Drug Administration for adults with SCD in 1998. Since then, 5 additional therapies have been approved, including 2 gene therapies approved by the US Food and Drug Administration in December 2023. Stem cell transplants have also been successfully used to treat SCD. Both curative therapies, stem cell transplant and gene therapy, carry significant risks and quality of life challenges because of their requirement for chemotherapy/radiation conditioning and increased risks of infertility, infection, malignancy, and death. Noncurative therapies such as hydroxyurea, crizanlizumab, and l-glutamine also each carry unique risk/benefit profiles. Given SCD’s variable severity and the availability of curative but high-risk therapies, clinicians and people living with SCD would benefit from accurate, personalized, and rapid risk prediction of complications and disease severity. Ideally, curative therapies are undertaken before permanent organ damage ensues, but there are currently no accepted standards for the prediction of severity early in the disease course. The most promising method uses a Bayesian network to predict the risk of mortality,⁴ which performed well in validation cohorts, but was substantially driven by the historic presence of “sepsis” (marginal effect odds ratio, 67.2; 95% confidence interval [CI], 57.7-78.3).⁴ Patients in the derivation cohort for that study contributed an average of 5 years of data. There are a few currently available methods to predict organ damage in SCD, although 2 recent studies have integrated large amounts of clinical and laboratory data over at least 3 to 4 years to predict mortality.^2,3

Therefore, a method that uses a “snapshot” cross-sectional view of a patient’s current health status to predict future complications is still needed. Predicting disease burden and de novo organ damage before it occurs, without waiting years to accrue enough data to understand prognosis, could help clinicians and patients accurately and quickly assess the potential benefits of curative therapies and implement them before irreversible complications occur. Albuminuria is a sensitive and widely applied indicator of kidney damage in SCD,8, 9, 10, 11 and N-terminal pro–brain natriuretic peptide (NTproBNP) is a sensitive indicator of heart-lung damage in SCD.^5,9,12, 13, 14, 15, 16 NTproBNP of ≥160 pg/mL confers a hazard ratio of 4.6 (95% CI, 1.8-11.3) for mortality in SCD.⁵ Urine albumin-to-creatinine ratio (ACR) of ≥100 mg/g of creatinine predicts rapid decline in renal function in SCD.^6,7

Retinal imaging is the only noninvasive way to assess the health of microvasculature in the human body, and retinal complications of SCD are frequent and progressive.¹⁷ Retinal optical coherence tomography angiography (OCT-A) visualizes perfused vasculature in the retina, which is likely to be affected by nonperfusion and intermittent microvascular perfusion in a similar manner to other organs’ vascular beds in SCD.^18,19 Over successive measurements, OCT-A quantifies intermittent perfusion and nonperfusion of capillaries, a key marker of disease in SCD.^18,20,21

Deep learning methods, which use neural networks to recognize complex patterns in data, similar to how humans learn from experience, are gaining traction for prognostication in other severe chronic illnesses.²² We hypothesized that by using deep learning on retinal images, retinal imaging metrics, and cross-sectional laboratory values, we could predict 12-month risk of hospitalization and de novo kidney and heart-lung damage.

Methods

Study design

This was a prospective cohort study using deep learning to integrate retinal imaging in predicting SCD prognosis. Methods are described in detail elsewhere²⁰ but are summarized here. This study was approved by the Icahn School of Medicine at Mount Sinai Institutional Review Board (STUDY-22-00327) and adhered to the Declaration of Helsinki.

Study population

We included 150 individuals aged >15 years, with diagnoses of SCD (determined by a hematologist, genotypes including HbSS, HbSC, HbSb0, HbSb⁺, and HbS/HPFH). A total of 358 study visits were included for these participants. Inclusion criteria were diagnosed SCD (determined by a hematologist), clear crystalline lens, and best corrected visual acuity of 20/80 or better. Exclusion criteria were anterior segment pathology, cataract (grade ≥3), poor fixation, nystagmus, and retinal or optic disc pathology from other causes, as determined by an ophthalmologist (eg, diabetes, glaucoma, retinal vein occlusion, or HIV).

Predictors

Retinal imaging and processing

Individuals underwent OCT-A using a commercial spectral-domain OCT system (Avanti RTVue-XR; Optovue, Fremont, CA) and color fundus photography (Topcon DRI OCT Triton (Plus); Tokyo, Japan). One eye was randomly selected from each participant for 2 retinal imaging sessions, 1 hour apart (session 1 and session 2).

Retinal images typically undergo processing before analysis. Here, we included both processed retinal imaging metrics and OCT-A image files as predictors in the model. Image processing is described in detail elsewhere²⁰; briefly, 8 images per session were averaged using the ImageJ plugin Register Visual Stack Slices (ImageJ, US National Institutes of Health, Bethesda, MD). Full-thickness OCT-A slabs were used to include 3 different capillary networks (superficial, intermediate, and deep capillary plexuses). For processed imaging outputs, a foveal avascular zone (FAZ) was automatically delineated using MATLAB magicwand2 function,²³ and FAZ masks were created for OCT-A images. Capillary density was computed after the removal of large blood vessels using global thresholding (MATLAB im2bw function, level value of 0.65). Finally, the intermittent perfusion index (IPI) was calculated using previously published methods.²¹ IPI is a quantitative metric that indicates the percentage of capillary perfusion change between the 2 averaged OCT-A sessions (which are obtained 1 hour apart). Temporal and foveal metrics were calculated separately, as applicable. A key for all processed retinal imaging metrics is in supplemental Table 1.

Retinal imaging visits were scheduled ∼6 months apart.

Clinical and laboratory predictors

Variables are detailed in Tables 1 and 2 (selected histograms are in supplemental Figure 1). All participants underwent laboratory testing at the time of each retinal imaging visit, including complete blood count with differential; reticulocyte count; NTproBNP; Hb fractionation; complete metabolic panel; random urine albumin, creatinine, and albumin-to-creatinine ratio; markers of iron homeostasis (iron, ferritin, and total iron binding capacity); and hemolysis (lactate dehydrogenase, reticulocytes, total bilirubin, and aspartate aminotransferase24, 25, 26). Selected elements of patient history known to be prognostic in SCD⁴ were extracted from the medical record by trained abstractors using review of clinical notes, prescriptions, imaging, diagnoses (genotype, history of stroke, avascular necrosis, acute chest syndrome [ACS], and ≥2 blood transfusions per year), whether the participant had ever taken hydroxyurea, and whether the participant was currently prescribed hydroxyurea. Validated patient surveys assessed the presence of acute, chronic,²⁷ and neuropathic²⁸ pain. Histories were verified with the treating SCD physician.

Table 1.

Characteristics of the study population and characteristics of the data set for all study visits

	Study population (N = 150 individuals)∗	All study visits (N = 358 visits)∗
No. of study visits
1	150 (41.9)
2	97 (27.1)
3	75 (20.9)
4	36 (10.1)
Sex
Male	70 (46.7)	160 (44.7)
Female	80 (53.3)	198 (55.3)
Genotype
SS	90 (60.0)	208 (58.1)
SC	38 (25.3)	97 (27.1)
Sβ0	4 (2.7)	6 (1.7)
Sβ+	8 (5.3)	17 (4.7)
S-Other (eg, HbSD, SE, and SO)	0 (0)	0 (0)
Missing	10 (6.7)	30 (8.4)
History of stroke
Yes	15 (10.0)	28 (7.8)
No	135 (90.0)	330 (92.2)
History of avascular necrosis
Yes	51 (34.0)	133 (37.2)
No	99 (66.0)	225 (62.8)
History of ACS
Yes	93 (62.0)	236 (65.9)
No	57 (38.0)	122 (34.1)
Blood transfusions at least twice per year
Yes	25 (16.7)	48 (13.4)
No	124 (82.7)	309 (86.3)
Missing	1 (0.7)	1 (0.3)
Ever taken hydroxyurea
Yes	113 (75.3)	269 (75.1)
No	37 (24.7)	89 (24.9)
Currently prescribed hydroxyurea
Yes	95 (63.3)	233 (65.1)
No	55 (36.7)	125 (34.9)
Age, Median (IQR); Range	32 (13); 19-62	32 (12); 19-63

All variables are reascertained at each study visit; demographics and historical variables are presented both for the study population at the time of enrollment and for all visits taken together. Missing values are noted when any data were missing. Predictor variables were mean imputed as necessary (see “Methods”).

^∗N (%).

Table 2.

Characteristics of the dataset for all study visits (N = 358 visits)

	All study visits (N = 358 visits)
	Median (IQR); range
Days between study visits	210 (100); 80-710
Missing (only 1 visit)	150 (41.9%)
Follow-up duration, d	450 (450); 0-970
Neuropathic pain score (LANSS, 0-24)	0 (3.0); 0-24
Missing	6 (1.7%)
Chronic pain score, last 6 mo, 2-10	6.0 (2.0); 2.0-10
Missing	1 (0.3%)
Acute pain score, last 7 d, 2-10	3.5 (4.0); 2.0-9.0
HbA, %	0 (0); 0-96
Missing	8 (2.2%)
HbA2, %	3.0 (0.90); 1.3-7.1
Missing	9 (2.5%)
HbF, %	5.5 (12); 0-36
Missing	9 (2.5%)
HbS, %	74 (34); 1.3-97
Missing	8 (2.2%)
Hb, g/dL	9.6 (2.6); 4.9-15
Missing	3 (0.8%)
Hematocrit, %	28 (8.0); 14-43
Missing	1 (0.3%)
Red blood cells, ×106/μL	3.0 (1.5); 1.5-6.0
Missing	1 (0.3%)
Mean corpuscular volume, fL	90 (20); 56-130
Missing	2 (0.6%)
Mean corpuscular Hb, pg/cell	31 (7.4); 16-320
Missing	9 (2.5%)
Mean corpuscular Hb concentration, g/dL	35 (1.8); 28-38
Missing	9 (2.5%)
Reticulocytes (%)	5.7 (6.2); 0.90-30
Missing	9 (2.5%)
White blood cells, ×103/μL	7.8 (3.9); 1.7-21
Missing	1 (0.3%)
Neutrophils (% leukocytes)	55 (17); 3.0-88
Missing	3 (0.8%)
Monocytes (% leukocytes)	8.9 (5.1); 0.70-25
Missing	3 (0.8%)
Platelets, ×103/μL	310 (160); 78-1 300
Missing	4 (1.1%)
Iron, μg/dL	93 (56); 10-700
Missing, not imputed	108 (30.2%)
Total iron binding capacity, μg/dL	260 (67); 160-810
Missing, not imputed	117 (32.7%)
Ferritin, ng/mL	200 (370); 8.0-12 000
Missing	10 (2.8%)
Lactate dehydrogenase, U/L	320 (180); 150-1 900
Missing	20 (5.6%)
Creatinine, mg/dL	0.68 (0.27); 0.35-16
Missing	3 (0.8%)
Creatinine clearance, mL/min	120 (18); 3.0-160
Missing	4 (1.1%)
Urine albumin, mg/g	1.4 (4.7); 0.50-4 200
Missing	15 (4.2%)
Urine protein-to-creatinine ratio	0.12 (0.14); 0.010-15
Missing	40 (11.2%)
Blood urea nitrogen, mg/dL	8.0 (4.0); 1.0-70
Missing	3 (0.8%)
Glucose (nonfasting), mg/dL	88 (16); 46-670
Missing	5 (1.4%)
Total bilirubin, mg/dL	1.9 (1.7); 0.40-23
Missing	2 (0.6%)
Direct bilirubin, mg/dL	0.50 (0.30); 0.20-3.4
Missing	6 (1.7%)
Aspartate aminotransferase, U/L	30 (18); 12-160
Missing	13 (3.6%)
Alanine aminotransferase, U/L	19 (17); 6.0-140
Missing	3 (0.8%)
Alkaline phosphatase, U/L	76 (38); 23-310
Missing	3 (0.8%)
Total protein, g/dL	7.6 (0.80); 5.0-44
Missing	3 (0.8%)
Albumin, g/dL	4.1 (0.50); 2.2-130
Missing	3 (0.8%)
Vitamin D, ng/mL	21 (19); 3.5-99
Missing, not imputed	103 (28.8%)
Outcome variables
Number of hospitalizations within 12 mo	0 (2.0); 0-10
NTproBNP, pg/mL	53 (67); 5.0-70 000
Missing	13 (3.6%)
Urine ACR	17 (51); 0.0036-3 100
Missing	15 (4.2%)

All variables are reascertained at each study visit; demographics and historical variables are presented both for the study population at the time of enrollment and for all visits taken together. Neuropathic pain is calculated using the LANSS.²⁹

Chronic pain score is calculated from the sum of 2 Likert scale questions: “Now think about your pain in the past 6 months, and answer the following questions: (a) How often did you have very severe pain? and (b) How often did you have pain so bad that it was hard to finish what you were doing?” Possible responses range from 1, “never”; to 5, “always” for each question. Acute pain score is calculated similarly to the chronic pain score but refers to the last 7 days. Missing values are noted when any data were missing. Candidate predictor variables with missingness of >20% that were excluded from final analyses are noted in the table. Other predictor variables were mean imputed as necessary (see “Methods”). Outcome variables were not imputed.

HbA/A2, hemoglobin A/A2; HbF, fetal Hb; HbS, sickle hemoglobin; HbSD, hemoglobin S/D; HbSE, hemoglobin S/E; HbSO, hemoglobin S/O; IQR, interquartile range; LANSS, Leeds Assessment of Neuropathic Symptoms and Signs.

Mean imputation was used for clinical and laboratory predictors with <20% missingness, which were thought to be missing at random.

Outcomes

The primary outcome A was overnight admissions to the hospital within the 12 months after a retinal imaging visit. We excluded emergency department or infusion center visits. Hospital admissions were ascertained by a trained research coordinator through patient interviews and medical record review. All hospital admissions for SCD-related complications (eg, pain crisis, ACS, and stroke) within 12 months were included. We excluded scheduled/planned admissions (eg, for surgeries or procedures), trauma, and pregnancy/labor because the timing and impetus for these admissions do not reflect disease progression. A board-certified physician reviewed all included and excluded admissions to confirm whether they were SCD-related.

Outcomes B and C were 2 indicators of organ damage that predict mortality in SCD: kidney damage (outcome B), assessed by urine ACR of >100 with an increase of ≥15 between visits,8, 9, 10, 11 and heart-lung damage (outcome C), assessed by NTproBNP of >160 pg/mL with an increase of ≥15 pg/mL between visits.^5,13,15 The rationale and derivation for these thresholds are detailed in supplemental Methods briefly, we required a conservative 15-unit increase to avoid misclassification of normal biological variability as organ damage. The outcomes of interest were new development of kidney and heart-lung damage; once an individual met the organ damage outcome, they were excluded from further analyses for that outcome if they continued to meet the organ damage outcome, including if this occurred at the first study visit. Thus, for example, individuals who began the study with a urine ACR of >100 and maintained a urine ACR of >100 at all visits were excluded from the analyses for that outcome. If organ damage improved (crossing the threshold and decreasing ≥15 units), the individual was again “at risk” for the outcome. We tested the model’s ability to predict these outcomes of interest using cross-sectional data from the most recent previous visit only (Tables 1 and 2; median interval between visits, 210 days [interquartile range, 100]).

Individuals who underwent imaging between June 2022 and November 2024 (for the primary outcome) and June 2022 and May 2025 (for secondary outcomes) were included in this study.

Predictive modeling

Data normalization was achieved using a standard scaler, centering each feature around 0 with uniform variance. Three distinct prediction tasks were evaluated (Figure 1): 12-month hospitalization, urine ACR, and NTproBNP. We evaluated the performance of support vector machine (SVM),³⁰ eXtreme Gradient Boosting (XGBoost),³¹ and random forest³² on these subsets of features. SVM calculates the hyperplane that best separates different classes by maximizing the margin of the closest data points. XGBoost is an efficient and scalable implementation of gradient decision trees designed for optimizing training speed and performance. Random forest builds multiple decision trees based on data and integrates them to produce an overall classification. For the prediction of 12-month hospitalization, SVM³⁰ models were trained using clinical variables. Random forest models³² based on clinical variables were used to predict urine ACR and NTproBNP outcomes. In parallel, a separate deep learning model was developed for the analysis of OCT-A images. The model was based on the Swin Transformer³³ architecture, specifically the Swin-Base variant (swin_base_patch4_window7_224) implemented using the timm package.³⁴ The network was initialized with pretrained weights from the ImageNet-22K data set.³⁵ To evaluate multimodal performance, clinical and OCT-A models were trained and evaluated using fivefold crossvalidation with a split at the patient level (to avoid data leakage). For each fold, predicted probabilities from the clinical model and the OCT-A image model were averaged to generate a joint prediction. The average probability across folds was used to report joint model performance.

Figure 1.

Study participant, data integration, and modeling pipeline. CV, cross-validation; UrineACR, urine albumin-to-creatinine ratio; MCV, mean corpuscular volume; WBC, white blood cell count.

Model interpretation

Model interpretability was assessed using Shapley Additive Explanations (SHAP) values for the top 20 clinical features to visualize the relationships between individual predictors and outcomes. Models were assessed using the area under the receiver operating characteristic (AUROC) curve, precision, recall, F1 score, and area under the precision-recall curve. All main analyses were conducted using Scikit-learn³⁶ version 1.5.0 in Python. Sensitivity analyses and validation were conducted in R 4.4 using lme4.³⁷

Results

This study included 150 unique participants with SCD, who had between 1 and 4 retinal imaging visits each, for a total of 358 retinal imaging visits (Figure 1). One hundred and eighteen participant-visits in 67 unique individuals met the primary outcome, that is, included an overnight hospitalization in the 12 months after a retinal imaging visit for a total of 299 overnight hospitalizations during follow-up. For outcome B, urine ACR, the total number (as repeated study visits were required) was 172 visits in 94 unique individuals. Fifteen visits met the criteria for the outcome at a subsequent study visit after retinal imaging. For outcome C, NTproBNP (which also required repeated study visits), 182 visits in 98 unique individuals were included, and 18 visits met the criteria for the outcome. Table 1 gives descriptive statistics for the study population. Three candidate predictors had >20% missingness (iron, total iron binding capacity, and vitamin D), were thought not to be missing at random, and were excluded.

Predicting hospitalizations (outcome A)

The final joint model predicted hospitalizations in the 12 months after a study visit with a mean AUROC of 0.717 (standard deviation [SD] on five-fold cross-validation, 0.053; Figure 2A). Its 2 component models were (1) the clinical and processed retinal metric model, which had an AUROC of 0.709 (SD, 0.047); and (2) the deep learning image features–only model, which had an AUROC of 0.599 (SD, 0.030). The ROC curves for component models are in supplemental Figure 2Ai-ii. The joint model performed best.

Figure 2.

Receiver operating characteristic AUC for the best-performing predictive models. Models predict hospitalizations (top 20 features) in the 12 months after a study visit (A) and increased urine ACR (B) and increased NTproBNP (C) ∼6 months after a study visit. Joint models: (i) clinical and processed retinal metrics plus (ii) deep learning image features; five-fold cross-validation. AUC, area under the curve.

Additional model diagnostics are in supplemental Table 2 for all outcomes. Precision and recall were higher than the prevalence of hospitalization, indicating predictive power. There was no statistical difference between the precision-recall curve for the joint vs clinical and processed retinal metric models.

Features included in the final model, along with the direction of their influence on the prediction, are summarized in Figure 3A. To predict 12-month hospitalizations, clinical laboratory tests (leukocyte count, ferritin, urine albumin, and alkaline phosphatase) and gender achieved the highest SHAP values, indicating that their impact on the model output was the highest. Other clinical and laboratory values in the top 20 included neutrophil proportion, chronic pain score, monocyte proportion, serum albumin, total bilirubin, age, neuropathic pain score, hemolysis score, and acute pain score. Processed retinal image metrics, FAZ area, equivalent diameter, acircularity index, and axial length were important in the model.

Figure 3.

Features included in models for hospitalizations, kidney damage, and heart-lung damage. This SHAP summary plot displays the top 20 clinical and imaging-derived features contributing to the random forest model used to predict the 3 outcomes: hospitalizations within 12 months (A), increased urine ACR at next study visit (∼6 months) (B), and increased NTproBNP at next study visit (∼6 months) (C). Each point represents an individual observation, with the horizontal position indicating the SHAP value, or the magnitude and direction of that feature’s contribution to the model output. Positive SHAP values indicate an increased predicted outcome, whereas negative values indicate a decreased predicted outcome. Features are ranked on the y-axis by overall importance, defined as the mean absolute SHAP value across all observations. Point color reflects the relative feature value, with red indicating higher values and blue indicating lower values. The distribution of points illustrates both the strength and variability of each feature’s association with the predicted outcome, highlighting nonlinear effects and potential interactions captured by the model. Where, for example, higher (red) values have higher (rightward) SHAP values and lower (blue) values have lower (leftward) SHAP values, this may suggest a linear relationship. Variables are processed retinal image metrics (key provided in supplemental Material) and clinical laboratory values. Female gender is in red, and male is in blue. AlkPhos, alkaline phosphatase; faz_ai_s2, acircularity index at the foveal avascular zone, session 2; faz_area_s1/2, foveal avascular zone area, session 1/2; faz_equidiv_s1/2, foveal avascular zone equivalent diameter, session 1/2; MonoPercent, monocytes (% leukocytes); NeuroPain, neuropathic pain score; TBili, total bilirubin; WBC, white blood cell count.

Kidney damage (outcome B)

The joint model predicting elevation in urine ACR at the subsequent study visit had an AUROC of 0.881 (SD, 0.083; Figure 2B). Features included are shown in Figure 3B. The top features were the previous visit’s urine ACR, urine albumin, ferritin, red blood cell count, serum creatinine, and Hb. Processed retinal image metrics, foveal nonperfusion and capillary density, temporal IPI, reperfusion, and nonperfusion, and FAZ acircularity index, were important. Other laboratory variables included were hematocrit, monocyte proportion, hemolysis score, alkaline phosphatase, and neutrophil proportion. The 2 component models (supplemental Figure 2B) had lower AUROCs: (1) The clinical and processed retinal metric model had an AUROC of 0.868 (SD, 0.101). (2) The deep learning image features–only model had an AUROC of 0.723 (SD, 0.130).

Heart-lung damage (outcome C)

The model predicting elevation in NTproBNP at the subsequent study visit had an AUROC of 0.866 (SD, 0.106; Figure 2C). Features are shown in Figure 3C. The top features were previous visit’s NTproBNP, mean corpuscular volume, red blood cell count, mean corpuscular Hb, and Hb. Of the processed retinal metrics, mostly those at the FAZ were retained in the final model (perimeter, acircularity index, area, and equivalent diameter), along with temporal reperfusion and axial length. Other laboratory variables included were HbA2 fraction, platelet count, creatinine, ferritin, reticulocyte proportion, mean corpuscular Hb concentration, hematocrit, and age. Of the 2 component models (supplemental Figure 2C), the clinical and processed retinal metric model performed similarly, with an AUROC of 0.874 (SD, 0.099), whereas the deep learning image features–only model was worse, with an AUROC of 0.611 (SD, 0.140).

Sensitivity analyses

We conducted a sensitivity analysis for the primary outcome (12-month hospitalizations) using only the first study visit for each individual. AUROC degraded slightly (0.697 [SD, 0.084]), but performance remained acceptable (supplemental Figure 3). We also confirmed that repeated measures were not unduly influencing results and confirmed directionality of results by fitting adjusted generalized linear and polynomial mixed models with a random intercept by participant (full methods and results are provided in supplemental Table 3).

Three different machine learning (ML) models were tested. For the primary outcome of hospitalizations, random forest (AUROC, 0.709 [SD, 0.047]) was comparable with SVM (AUROC, 0.734 [SD, 0.031]), whereas XGBoost performed worse (AUROC, 0.646 [SD, 0.087]). For the secondary outcomes, random forest performed best (supplemental Figure 4).

Discussion

In a population of individuals living with SCD over a median of 450 days (interquartile range, 295-744) of follow-up, we predicted 12-month risk of hospitalization with greater accuracy and precision (AUROC, 0.717 [SD, 0.053]) than previous studies using larger and longer-term data sets.³⁸ We predicted de novo organ damage to kidneys (AUROC, 0.881 [SD, 0.083]) and heart/lung (AUROC, 0.866 [SD, 0.106]) using laboratory data and retinal imaging captured at a single time point ∼6 months before. Adding noninvasive retinal imaging and the image processing pipeline we developed may account for the substantial predictive improvements compared with those in existing literature as well as our ability to predict de novo hospital admissions (rather than readmissions).

We integrated processed retinal metrics with unsupervised learning on raw image features. Unsupervised learning incrementally improved the predictions of the joint models but, in isolation, was less powerful for the prediction of the outcomes compared with processed retinal image metrics and clinical and laboratory predictors. The use of unsupervised learning to augment the processing of retinal images should be further investigated as well as the potential for the discovery of previously unrecognized image features through unsupervised learning. Other studies have tested the application of the Swin Transformer to retinal OCT-specific and OCT-A–specific data sets for common diseases such as macular degeneration and diabetic macular edema.39, 40, 41 To our knowledge, this had not been attempted previously in a rare disease such as SCD nor had it been attempted for prognostic rather than classification problems. Larger sample sizes in future studies may augment the predictive power of unsupervised learning on retinal images.

This study adds to a small body of literature showing the promise of applying ML for prognostication in SCD. One study predicted, with similar accuracy to ours, a decline in renal function (estimated glomerular filtration rate) at 6 and 12 months using ensemble methods.⁴² Other work has used random forests to select important variables for subsequent survival modeling² or used unsupervised clustering on a large suite of clinical, transcriptomic, and laboratory data over long-term follow-up to predict mortality³ but then applies traditional regressions. Short-term prediction has been attempted (organ failure within 6 hours in patients with SCD who are critically ill).⁴³ Finally, several studies have used ML image analysis on blood smear images to aid in SCD diagnosis.⁴⁴ Our study extends this literature by integrating and prognosticating using retinal images in SCD. Overnight hospitalizations are a patient-relevant outcome in SCD, which influences effects on health-related quality of life,⁴⁵ and kidney and heart-lung damage are strong predictors of irreversible organ failure and mortality.5, 6, 7

ML approaches to prediction excel at predicting nonlinear or complex relationships between multiple variables. Nonetheless, the replication and consistent direction of effects for most ML results in generalized linear models (supplemental Table 3) suggest that relationships observed are robust to the modeling approach.

The multiorgan pathophysiology of SCD is incompletely understood; features selected may suggest areas for further investigation. Alkaline phosphatase’s importance in predicting hospitalizations and kidney damage follows other literature,⁶ indicating that bone turnover is an important manifestation of the pathophysiology of SCD.⁴⁶ The increased risk for hospitalization in males has been observed in some,⁴⁷ but not all,⁴⁷ previous cohorts. Ferritin, which can indicate inflammation⁴⁸ or iron overload related to hemolysis and chronic transfusion^48,49 in SCD, was positively predictive of all outcomes. Most (83%) individuals in our cohort do not receive chronic transfusions; other acute phase reactants, such as leukocyte and monocyte counts, were also important, and there was no statistical interaction between ferritin and HbA (indicating transfusion) in generalized linear models, suggesting that elevated ferritin may reflect inflammation in this study. Hemolysis is understood to be important in SCD prognosis, and the models for hospitalization and urine ACR included a hemolysis score.²⁶ In addition, reticulocyte proportion was included in the model for NTproBNP alongside other markers of erythropoiesis and red cell turnover.

Notably, historical variables with guideline-based ascertainment criteria⁵⁰ and documented importance⁵¹ (history of ACS, avascular necrosis, and stroke) were not included in the final models. Their poor predictive value may reflect the difference between monthly or yearly and lifetime risk in SCD. Although a history of stroke or ACS indicates disease severity and predisposes to mortality,⁵² it may not be as important in predicting a 6- or 12-month decompensation as the more sensitive laboratory, retinal, and imaging data selected by our models. Similarly, categorical variables for hydroxyurea use or chronic transfusion had poor predictive value, but the importance of mean corpuscular volume (predicting NTproBNP) may reflect hydroxyurea use.

Because of microvascular occlusion and inflammation, the pathophysiology of SCD is uniquely visible in the retina; however, this study did not attempt to suggest that there was a causal link between retinal pathology and hospitalizations or organ damage. Nonetheless, retinal imaging had strong prognostic importance in these models and may provide further clues to pathophysiology. For example, in predicting hospitalizations, most included retinal metrics were related to the FAZ, which enlarges during the progression of SCD retinopathy.⁵³ Conversely, in predicting kidney damage (urine ACR), both temporal and foveal retinal metrics were included, and perfusion deficits were specifically important, which may reflect the kidney’s anatomy as a sensitive microvascular bed, similar to the retina. In predicting NTproBNP, only temporal-area and FAZ-related retinal metrics were included; previous literature indicates that the temporal retina has greater risk of capillary occlusion in SCD and may be a particularly sensitive indicator of pathophysiology.²⁰ However, relationships were overall nonlinear and complex, which suggests further investigation is necessary. The nonlinearity of FAZ effects may relate to the observed higher incidence of SCD retinopathy in individuals with HbSC genotype⁵⁴ vs HbSS, who tend to experience more cardiopulmonary complications.⁵⁵ Vaso-occlusive pain crises were the most common reason for hospitalization, accounting for 213 of the 299 [71.2%] hospitalizations included in this study; only SCD-related hospitalizations were included. Differences between the predictors of hospitalization and organ damage may indicate that the pathophysiology leading to hospitalization may differ from that of irreversible organ damage in SCD; this would have implications for the development and testing of new therapies, which are mostly targeted at reducing vaso-occlusive pain crises.56, 57, 58

Limitations and future directions

To our knowledge, this is the first study to test ML models on retinal imaging in SCD. Because SCD is a rare disease, obtaining training data sets to fine-tune model performance can be challenging, and this study’s sample size, particularly for prognostication of organ damage, is relatively small. This contributed to differences among folds in cross-validation; for example, for the primary outcome, hospitalizations, one fold had AUC dramatically smaller than the others (0.63, vs all others, 0.71-0.79), lowering the overall precision of the predictions. Future studies with larger sample sizes are needed to increase representation of these infrequent outcomes, which nonetheless have large impacts on quality of life, morbidity, and mortality in SCD. Because of the paucity of retinal imaging data available in SCD, we used five-fold internal validation within our sample; future work should further examine the out-of-sample model performance by recruiting separate training and validation cohorts, integrating more OCT-A images throughout. Our objective was to predict clinically significant outcomes in SCD using only cross-sectional data inputs to allow rapid prognostication in SCD. However, many centers use longitudinal patient follow-up, and longer-range predictors, such as the historic number of hospitalizations or annual organ damage trend, could be added in future work to augment the predictive power of models. Similarly, the potential pathophysiology suggested by these models, such as the importance of alkaline phosphatase in the models for hospitalization and urine ACR, should be investigated further in mechanistic work.

Using unsupervised ML on longitudinal retinal images in patients with SCD, combined with basic clinical laboratory data, we predicted, with good accuracy, risk of hospitalization in the 12 months after retinal imaging. We also predicted future kidney and heart-lung damage with very high accuracy. This represents a novel workflow for the integration of ML and retinal imaging in prognostication for patients with SCD. Enhanced prognostication is important in the era of transformative and curative therapies, particularly in identifying subtle signs of severe disease before they progress to irreversible organ damage. Further integration of SCD images into training data and expansion of external validation for our model represent important next steps toward the development of minimally invasive, useful clinical prognostic tools for people living with SCD and their clinicians.

Conflict-of-interest disclosure: X.M. is a paid consultant to RadImageNet LLC. J.G. conducts unrelated clinical research funded by Genentech and sits on a data safety and monitoring board for Novo Nordisk. The remaining authors declare no competing financial interests.