Sie können Operatoren mit Ihrer Suchanfrage kombinieren, um diese noch präziser einzugrenzen. Klicken Sie auf den Suchoperator, um eine Erklärung seiner Funktionsweise anzuzeigen.
Findet Dokumente, in denen beide Begriffe in beliebiger Reihenfolge innerhalb von maximal n Worten zueinander stehen. Empfehlung: Wählen Sie zwischen 15 und 30 als maximale Wortanzahl (z.B. NEAR(hybrid, antrieb, 20)).
Findet Dokumente, in denen der Begriff in Wortvarianten vorkommt, wobei diese VOR, HINTER oder VOR und HINTER dem Suchbegriff anschließen können (z.B., leichtbau*, *leichtbau, *leichtbau*).
While many deep learning systems (DLSs) for diabetic retinopathy (DR) have been developed and validated on cohorts with an average age of 50s or older, fewer studies have examined younger individuals. This study aimed to understand DLS performance for younger individuals, who tend to display anatomic differences, such as prominent retinal sheen. This sheen can be mistaken for exudates or cotton wool spots, and potentially confound DLSs.
Methods
This was a prospective cross-sectional cohort study in a “Diabetes of young” clinic in India, enrolling 321 individuals between ages 18 and 45 (98.8% with type 1 diabetes). Participants had fundus photographs taken and the photos were adjudicated by experienced graders to obtain reference DR grades. We defined a younger cohort (age 18–25) and an older cohort (age 26–45) and examined differences in DLS performance between the two cohorts. The main outcome measures were sensitivity and specificity for DR.
Results
Eye-level sensitivity for moderate-or-worse DR was 97.6% [95% confidence interval (CI) 91.2, 98.2] for the younger cohort and 94.0% [88.8, 98.1] for the older cohort (p = 0.418 for difference). The specificity for moderate-or-worse DR significantly differed between the younger and older cohorts, 97.9% [95.9, 99.3] and 92.1% [87.6, 96.0], respectively (p = 0.008). Similar trends were observed for diabetic macular edema (DME); sensitivity was 79.0% [57.9, 93.6] for the younger cohort and 77.5% [60.8, 90.6] for the older cohort (p = 0.893), whereas specificity was 97.0% [94.5, 99.0] and 92.0% [88.2, 95.5] (p = 0.018). Retinal sheen presence (94% of images) was associated with DME presence (p < 0.0001). Image review suggested that sheen presence confounded reference DME status, increasing noise in the labels and depressing measured sensitivity. The gradability rate for both DR and DME was near-perfect (99% for both).
Conclusion
DLS-based DR screening performed well in younger individuals aged 18–25, with comparable sensitivity and higher specificity compared to individuals aged 26–45. Sheen presence in this cohort made identification of DME difficult for graders and depressed measured DLS sensitivity; additional studies incorporating optical coherence tomography may improve accuracy of measuring DLS DME sensitivity.
Prior Presentation: poster at The Association for Research in Vision and Ophthalmology (ARVO), May 5–9, 2024, Seattle, WA.
Key Summary Points
Validation of deep learning systems (DLSs) to detect diabetic retinopathy (DR) usually focus on patient cohorts in their 50s.
This study aimed to understand if DLSs for DR work well on younger patients, who tend to have retinal sheen.
The DLS had high sensitivity and specificity for moderate-or-worse DR. The DLS also had high specificity for diabetic macular edema (DME), but DME sensitivity was lower than expected. Comparing patients aged 18–25 versus 26–45, specificity was higher in the younger patients for both DR and DME.
On image inspection, lower DME sensitivity in this cohort could be due to difficulty in the reference grading in the presence of retinal sheen; additional research using additional imaging (such as optical coherence tomography) will be needed to confirm this.
DR detection worked well in this younger cohort, and did not appear to be affected by retinal sheen.
Introduction
Multiple studies have shown that deep learning systems (DLSs) can detect diabetic retinopathy (DR) from fundus images with high accuracy [1‐4]. However, fewer studies have focused on the performance of DR screening DLSs in a young adult population [5]. This population is of particular importance as the incidence of diabetes mellitus (DM) is rapidly increasing in the adolescent and young adult population, and the rates of DR are similarly expected to rise in this population [6‐8].
Anzeige
Model performance in young adults may be impacted by differences in fundus anatomy associated with younger age. Retinal sheen, more commonly found in younger individuals, can be mistaken for exudates or a cotton wool spot [9]. Other anatomical differences (e.g., lower rates of cataracts, corneal scars, vitreous opacities, etc.) may alter image gradability and DLS performance, particularly since the DLS has been trained using datasets from older populations [1].
In this prospective observational study, we investigate the performance of a DLS [1, 10, 11] for detecting DR in a young adult population with diabetes in New Delhi, India.
Methods
Individuals aged 18 to 45 attending the “Diabetes of Young” clinic at the All India Institute of Medical Sciences (AIIMS), New Delhi hospital were prospectively consented and recruited (Table 1). A formal sample size calculation was not conducted because of the unknown DLS performance on this population. The target sample size was selected qualitatively to be in the hundreds, incorporating expectations that a minority of patients overall would have DR or diabetic macular edema (DME), and to leave sufficient sample size to examine subgroup trends (such as age). Between April 2021 and August 2022, 343 individuals aged 18 to 45 were screened. Twenty-two (22) individuals were excluded because of a history of cataracts. Two cohorts were defined: a younger cohort (age 18–25) and an older cohort (age 26–45). The younger cohort consisted of 162 individuals (324 eyes) and the older cohort consisted of 159 individuals (318 eyes). All participants had both pupils dilated and underwent fundus photography using a Topcon Maestro camera. Images were processed by an automated gradability algorithm [11], and up to three photos were attempted before an image was deemed ungradable.
DR diabetic retinopathy, DME diabetic macula edema, NPDR non-proliferative DR, PDR proliferative DR, SD standard deviation
aHbA1c percentage not available for 8/162 younger participants and 7/159 older participants
This study was approved by the Institutional Ethics Committee at AIIMS (# IEC-913/03.01.2020, RP-23/2020), and all research followed the Declaration of Helsinki. All participants consented and their data were de-identified by AIIMS prior to analysis.
Anzeige
All fundus photographs were cropped to remove the background and resized to 779 × 779 pixels prior to input to the DLS [1, 10, 11]. The DLS was developed using over one million images (with an average age above 50) and the Inception-v4 architecture, and provided outputs for (1) DR severity and (2) DME presence. DR was classified following the International Clinical Diabetic Retinopathy (ICDR) scale as no DR, mild non-proliferative DR (NPDR), moderate NPDR, severe NPDR, or proliferative DR (PDR). DME was defined by the presence or absence of hard exudates within one disc diameter of the macula. A detailed breakdown is provided in Supplementary Table 5. The reference grades for each case were arrived at via adjudication [10] by three graders (two ophthalmologists and one optometrist), with the majority opinion taken if consensus was not reached within three rounds of discussions [12].
The primary analysis was a comparison of sensitivity and specificity for moderate or worse DR and DME presence between the younger and older cohort. Secondary analyses comparing the younger and older cohort included (1) PDR sensitivity and specificity, (2) vision-threatening DR (VTDR) sensitivity and specificity, and (3) DR and DME gradability. Confidence intervals were calculated using bootstrapping, and hypothesis testing used the permutation test [13], both at the patient level to account for correlation across eyes.
To better understand the impact of retinal sheen on model performance, we performed two-tailed Fisher exact tests for DR/DME reference grades and sheen presence after two ophthalmologists further evaluated fundus photos for confounding effects of sheen. Sheen is defined as contiguous shiny or reflective areas most commonly around the fovea and retinal vasculature, and is considered a feature of the retina surface and is distinct from intraretinal (i.e., exudates) or subretinal (i.e., drusen) abnormalities. The presence of sheen was annotated visually on the basis of its characteristic appearance without a requirement for area or size.
To further identify which factors were correlated with the diagnostic accuracy of the DLS, we performed a multivariable logistic regression analysis using age, sex, HbA1c percentage, duration of diabetes, and sheen presence as features. We performed this analysis on both sensitivity and specificity for moderate+ DR (i.e., moderate, severe, or proliferative DR) and DME. Age features were bucketed with the same younger/older cohort cutoff, and HbA1c percentage and duration of diabetes were bucketed to achieve roughly equal-sized buckets (Supplementary Table 2).
Results
Patients predominantly had type 1 diabetes (> 98% in both younger and older groups, Table 1). For 642 eyes (321 individuals; see Supplementary Fig. 1), eye-level sensitivity and specificity for moderate+ DR were 95.1% [95% CI 91.0, 97.8] and 95.3% [93.2, 97.2], respectively. Sensitivity trended higher for the younger group (97.9% [95.9, 99.3]) compared with the older group (92.1% [87.6, 96.0]); p = 0.418. The DLS had higher specificity for the younger group (97.8% [96.0–99.3]) compared to the older group (92.1% [87.6–96.0]); p = 0.008 (Table 2).
Table 2
Eye-level model performance
Endpoint
Metric
Full cohort (ages 18–45), N = 642
Younger cohort (ages 18–25), N = 324
Older cohort (ages 26–45), N = 318
P (younger vs. older cohort)
Moderate-or-worse DR
Sensitivity
95.1% [91.0, 97.8]
97.6% [91.2, 98.2]
94.0% [88.8, 98.1]
0.418
Specificity
95.3% [93.2, 97.2]
97.9% [95.9, 99.3]
92.1% [87.6, 96.0]
0.008
DME
Sensitivity
78.0% [65.0, 87.9]
79.0% [57.9, 93.6]
77.5% [60.8, 90.6]
0.893
Specificity
94.6% [92.3, 96.6]
97.0% [94.5, 99.0]
92.0% [88.2, 95.5]
0.018
VTDRa
Clinically important miss rate
7.5% [2.22, 14.29]
4.0% [0.00, 13.33]
9.1% [2.17, 17.54]
0.002
mtmDRb
Sensitivity
95.1% [91.3, 98.5]
97.6% [92.50, 100]
94.1% [89.26, 98.23]
0.330
Specificity
95.1% [92.9, 97.2]
97.9% [95.99, 99.30]
91.6% [87.32, 95.54]
0.008
Gradability rate of DLS
DR
99.2% [98.2, 99.7]
99.1% [97.3, 99.7]
99.4% [97.7, 99.8]
1.00
DME
99.1% [98.0, 99.6]
99.1% [97.3, 99.7]
99.1% [97.3, 99.7]
1.00
Patient-level performance is reported in Supplementary Table 3
DR diabetic retinopathy, DME diabetic macula edema, VTDR vision-threatening DR, mtmDR more-than-mild DR, DLS deep learning system
aVTDR is defined as severe NPDR, proliferative DR (PDR), or DME
bmtmDR is defined as moderate or severe NPDR, proliferative DR, or DME
Sensitivity for DME was 78.0% [65.0, 87.9] overall and not significantly different (p = 0.893) between the younger (79.0% [57.9, 93.6] and older (77.5% [60.8, 90.6]) cohorts. DME specificity was 94.6% [92.3, 96.6] overall and was higher for younger (97.0% [94.5, 99.0]) than older 92.0% [88.2, 95.5] groups (p = 0.018) (Table 2). Retinal sheen was observed in 94% of images and included all false negatives (7 images for DR, 13 images for DME; Fig. 1). Sheen presence was associated with DME presence based on the reference grades (p < 0.0001) (Supplementary Table 4).
Fig. 1
False negative examples. Left: false negative diabetic macula edema (DME) examples, where the images were adjudicated by experienced graders as DME positive, but the deep learning system (DLS) did not flag DME. Right: false negative moderate+ diabetic retinopathy (DR) examples, where the images were adjudicated as moderate+ DR positive, but the DLS did not flag moderate+ DR. All images, on both the left and right, are positive for sheen; the sheen is more visually evident on the left images. Across all images, white arrows indicate regions that appear to be hard exudates; with a magnified crop shown in the top left corner
In multivariable analysis across the entire dataset, age group and sheen presence were significantly (p < 0.05) associated with model errors for both DR and DME true negatives (Supplementary Table 2). In other words, DLS specificity was associated with both age and sheen presence.
The DLS’s overall DR gradability rate was 99.2%. There was no significant difference between younger (99.1%) and older (99.4%) groups (p = 1.0). Gradability for DME was also excellent with near identical performance (99.1%) between younger and older groups (Table 2).
Anzeige
Discussion
This cross-sectional study validates the high performance of our DLS screening for DR and DME in a younger population with predominantly type 1 diabetes. This patient population is important to understand both as a result of their rates of diabetes-associated complications [6‐8] and complex factors influencing DR screening compliance [14, 15].
We identified four key findings: (1) sensitivity and specificity for detecting moderate+ DR were both 95%; (2) sensitivity and specificity for detecting moderate+, VTDR, and PDR were similar or better in the younger group, although the differences for sensitivity were not statistically significant; (3) sensitivity and specificity for detecting DME were 78% and 95%, respectively, with all misses occurring in images with retinal sheen; (4) image gradability was near perfect (99% for both DR and DME). Despite observed specificity differences between the younger and older groups, specificity was high at 92–98% for both DR and DME, and both age groups. These findings confirm the DLS performs well in a young adult population, with the caveat that DME sensitivity was lower than in prior studies.
The relatively lower-than-expected DME sensitivity (independent of age group) was investigated by review of the DME images with two ophthalmologists. Retinal sheen, which is more prevalent in younger populations, can be misidentified as exudates in 2D fundus photography. When retinal sheen is isolated to the macula or is noncontinuous, mimicking the appearance of individual exudates, it can be challenging to distinguish between the two findings (Fig. 1). As a result, the graders may have erred towards conservatively overcalling DME, resulting in the DLS’s apparent undercalling and thus relatively lower DME sensitivity. Future work may aid in determining reference DME status using modalities invariant to sheen, such as optical coherence tomography (OCT).
The higher specificity for DR and DME in the younger cohort relative to the older cohort in this study (Table 2) runs counter to our original hypothesis of possible DLS lesion recognition errors in the presence of sheen. It is possible that the above reasoning related to challenging exudate determination played a nuanced role, by shifting these challenging negative images to the positive set, and doing so more often in the younger cohort where sheen was more frequently present. Assuming that on average these challenging negative images were often called positive by the DLS, this effectively shifts counts from the false positives (DLS predicted positive but ground truth negative) to the true positives (DLS predicted positive and ground true positive). This results in relative increases in both sensitivity and specificity, the trend observed. As above, these trends will need to be confirmed using OCT, though a reasonable conclusion from the current data is that younger age groups and sheen presence do not appear to result in decreased performance.
Anzeige
Beyond DR and DME performance, DLS gradability rates observed in this study were notably high (99% for DR and DME). This is substantially higher than previously reported gradability rates between 71% and 91% [16]. In this study, all individuals were relatively young and underwent pupil dilation which contributes to the high gradability rate.
This study’s limitations include a predominantly single-ethnicity population in India, and the requirement of pupillary dilation prior to image acquisition. Further studies with diverse populations and the addition of multi-field or stereoscopic fundus photography, and OCT imaging could improve generalizability of the findings and precision of DR/DME findings. It would also be useful in future work to directly compare a younger patient cohort with individuals older than 45. Finally, as a result of the correlation between age and sheen presence, it is challenging to disentangle the association of DLS performance with age, versus with sheen.
Conclusion
In this prospective study, we investigated the performance of our DLS on a young adult patient population with predominantly type 1 diabetes. The DLS’s sensitivity and specificity were high for DR, and the relatively lower DME sensitivity could be explained by confounding effects of retinal sheen and will require further investigation incorporating OCT.
Acknowledgements
We thank the participants of the study.
Anzeige
Declarations
Conflict of Interest
Antonio Tan-Torres III, Divleen Jeji, Xiang Yin, Lu Yang, Preeti Singh, Tayyeba Ali, Dushyantsinh Jadeja, Rajroshan Sawhney, Dale R Webster, Naama Hammel, Yun Liu, Kasumi Widner, Sunny Virmani, Jonathan Krause are current or past employees of Google or an affiliate and own Alphabet stock. Arthur Brant, Ilana Traynis are paid consultants of Google. Pradeep A Praveen, Pradeep Venkatesh and Nikhil Tandon have nothing to disclose.
Ethical Approval
This study was approved by the Institutional Ethics Committee at AIIMS (# IEC-913/03.01.2020, RP-23/2020), and all research followed the Declaration of Helsinki. All participants consented and their data were de-identified by AIIMS prior to analysis.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License, which permits any non-commercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc/4.0/.
Validation of a Deep Learning Model for Diabetic Retinopathy on Patients with Young-Onset Diabetes
Verfasst von
Antonio Tan-Torres III
Pradeep A. Praveen
Divleen Jeji
Arthur Brant
Xiang Yin
Lu Yang
Preeti Singh
Tayyeba Ali
Ilana Traynis
Dushyantsinh Jadeja
Rajroshan Sawhney
Dale R. Webster
Naama Hammel
Yun Liu
Kasumi Widner
Sunny Virmani
Pradeep Venkatesh
Jonathan Krause
Nikhil Tandon
Gulshan V, Peng L, Coram M, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA. 2016;316:2402–10.CrossRefPubMed
2.
Ting DSW, Cheung CY-L, Lim G, et al. Development and validation of a deep learning system for diabetic retinopathy and related eye diseases using retinal images from multiethnic populations with diabetes. JAMA. 2017;318:2211–23.CrossRefPubMedPubMedCentral
3.
Abràmoff MD, Lavin PT, Birch M, et al. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. NPJ Digit Med. 2018;1:39.CrossRefPubMedPubMedCentral
4.
Ipp E, Liljenquist D, Bode B, et al. Pivotal evaluation of an artificial intelligence system for autonomous detection of referrable and vision-threatening diabetic retinopathy. JAMA Netw Open. 2021;4:e2134254.CrossRefPubMedPubMedCentral
5.
Lo J-E, Kang EY-C, Chen Y-N, et al. Data homogeneity effect in deep learning-based prediction of type 1 diabetic retinopathy. J Diabetes Res. 2021;2021:2751695.CrossRefPubMedPubMedCentral
6.
Jensen ET, Rigdon J, Rezaei KA, et al. Prevalence, progression, and modifiable risk factors for diabetic retinopathy in youth and young adults with youth-onset type 1 and type 2 diabetes: the SEARCH for diabetes in youth study. Diabetes Care. 2023;46:1252–60.CrossRefPubMedPubMedCentral
7.
TODAY Study Group. Development and progression of diabetic retinopathy in adolescents and young adults with type 2 diabetes: results from the TODAY study. Diabetes Care. 2021;45:1049–55.CrossRefPubMedCentral
8.
Cho Y, Park H-S, Huh BW, et al. Prevalence and risk of diabetic complications in young-onset versus late-onset type 2 diabetes mellitus. Diabetes Metab. 2022;48:101389.CrossRefPubMed
9.
Saha SK, Xiao D, Kanagasingam Y. A novel method for classification of exudates and retinal sheen in fundus photographs for automated disease grading. Invest Ophthalmol Vis Sci. 2018;59:1694–1694.
10.
Krause J, Gulshan V, Rahimy E, et al. Grader variability and the importance of reference standards for evaluating machine learning models for diabetic retinopathy. Ophthalmology. 2018;125:1264–72.CrossRefPubMed
Phene S, Dunn RC, Hammel N, et al. Deep learning and glaucoma specialists: the relative importance of optic disc features to predict glaucoma referral in fundus photographs. Ophthalmology. 2019;126:1627–39.CrossRefPubMed
13.
Chihara LM, Hesterberg TC. Mathematical statistics with resampling and R. Hoboken: Wiley; 2022.
14.
Curran K, Ahmed M, Sultana MM, et al. Adherence to diabetic retinopathy screening among children and young adults in Bangladesh. Clin Diabetes Endocrinol. 2024;10:41.CrossRefPubMedPubMedCentral
15.
Prothero L, Cartwright M, Lorencatto F, et al. Barriers and enablers to diabetic retinopathy screening: a cross-sectional survey of young adults with type 1 and type 2 diabetes in the UK. BMJ Open Diabetes Res Care. 2022. https://doi.org/10.1136/bmjdrc-2022-002971.CrossRefPubMedPubMedCentral
16.
Raman R, Rani PK, Mahajan S, et al. The tele-screening model for diabetic retinopathy: evaluating the influence of mydriasis on the gradability of a single-field 45 degrees digital fundus image. Telemed J E Health. 2007;13:597–602.CrossRefPubMed
Nicht nur Schlaganfälle, sondern auch systemische embolische Ereignisse (SEE) stellen für Menschen mit Vorhofflimmern eine Gefahr dar, wie eine Metaanalyse deutlich macht. Schutz bieten vor allem direkt wirksame orale Antikoagulanzien (DOAK).
Das experimentelle Antikoagulans Abelacimab hat gegenüber Rivaroxaban den Vorteil eines geringeren Blutungsrisikos. Ältere Menschen könnten davon besonders profitieren.
In einer Autopsiestudie aus Japan fand sich in einem substanziellen Anteil der Fälle eine zu Lebzeiten nicht diagnostizierte Krebserkrankung, häufig ein wenig aggressives Prostata- oder Schilddrüsenkarzinom. Einige wenige Tumoren hatten allerdings unbemerkt gestreut.
Ob und wie stark können Infektionen mit dem Respiratorischen Synzytial-Virus kardiorespiratorische Probleme auch in der postakuten Phase verursachen? Eine Studie hat das untersucht.
