Using Deep Learning Models to Predict the Risk of Peripheral Neuropathy on Diabetic Patients

doi:10.21203/rs.3.rs-2835649/v1

To establish a suitable deep learning model in order to predict the risk of diabetic Peripheral Neuropathy (DPN) by using fundus photography images of type II diabetic patients with artificial intelligence (AI) approach. From the year 2013 to 2017, by using a diabetes care database established by the Department of Endocrinology and Metabolism of Kaohsiung Datong Hospital (KMTTH) and the Affiliated Hospital of Kaohsiung Medical University (KMUH), a corresponding retrospective study of fundus photographic images of patients with type II diabetes mellitus was performed. Patients with Type II diabetes who have undergone clinical routine treatment and have routine ophthalmoscopy were analyzed and classified according to the results of the Nerve Conduction Velocity (NCV) method. Soon after the patient's personal information was removed, the image is preprocessed by adaptive histogram equalization (CLAHE) to limit the contrast variation. These preprocessed images were then divided into training, validation, and test sets. Another two sets of Rotated image data were also incorporated for enhancements to build prediction models through four deep learning architectures: InceptionNet, VGGNet, ResNet, and ConvMixer DPN model, respectively. In this study, a classification model for predicting the severity of DPN was successfully established through four deep learning architectures. The accuracies of the four DPN prediction models established in this study were 0.94, 0.90, 0.97, and 0.96; AUC values have achieved at 0.92, 0.93, 0.95, and 0.96; specificity analyses were 1.00, 0.92, 1.00, and 0.98; The combined sensitivity values of mild and moderate to severe DPN reached 0.84, 0.90, 0.90, and 0.92, respectively. An AI-assisted diagnostic model was successfully established to predict the severity of diabetic peripheral neuropathy (DPN), which could determine whether the patient has DPN from the retinal fundus images obtained after ophthalmoscopy and with its associated severity; therefore, it is an efficient, non-invasive method of DPN detection.

Health sciences/Diseases/Eye diseases

Health sciences/Diseases

Physical sciences/Engineering

retinal fundus imaging

diabetic peripheral neuropathy

type II diabetes

nerve conduction examination

deep learning

Diabetic peripheral neuropathy (DPN) is a common complication of diabetes mellitus (DM) that can lead to foot ulcers or even the need for amputation. Therefore, how to screen and manage diabetic peripheral neuropathy (DPN) in the early stage of the disease is an important goal for many researchers and clinicians. Symptoms of DPN usually become worse over a certain period as the disease progresses, and patients may experience signs such as numbness when wearing socks or gloves, or even losing ankle reflex [1, 2] which tends to worsen at night. These early symptoms can affect the quality of the patient's sleep, and may even cause long-term fatigue and tiredness [3].

Clinically, diabetic patients may often experience burning and needle-sticking pain in extremities; some may encounter numbness and/or other symptoms by self-observation. Patients usually have begun to bear these chronic pains which may obstruct daily lives. Therefore, how to diagnose the early disease stages of DPN is one of the critical steps to reducing the side effects as the disease progresses. Certain wounds of diabetic patients are not easy to heal due to problems such as hyperlipidemia, the most commonly utilized non-invasive examination method for neuropathy is Nerve Conduction Velocity (NCV) examination; however, the disadvantage of NCV in early diabetic stages is that it lacks sufficient sensitivity to identify lesions consisting of small nerve fibers [4–6].

Artificial intelligence (AI) has been widely adopted in the medical field in recent years. Typical applications include pixel-level object detection for digital pathological slice images of various cancers, and lung CT imaging detection for viral pneumonia [7, 8], which have confirmed the feasibility and practicality of artificial intelligence advantages applied to medical imaging research and application.

In ophthalmology, color retinal imaging is being widely utilized in the screening of diabetic retinopathy (DR). According to past and ongoing research, chronic vascular complications of diabetes such as retinopathy have a significant correlation with peripheral neuropathy in patients with Type II diabetes mellitus [9, 10]. The purpose of this study is to investigate the potential strong correlation between retinal fundus imaging and diabetic peripheral neuropathy through the establishment of various deep learning models. This research hopes to detect the diabetic fundus images automatically and efficiently through AI models in order to determine whether patients have developed DPN and/or its corresponding severity level for further disease management concerns.

In this designed study framework, a diabetes care database was established by collecting fundus photographic images of patients with Type II diabetes retrospectively from the Department of Endocrinology and Metabolism of Kaohsiung Datong Hospital (KMTTH) and the Affiliated Hospital of Kaohsiung Medical University (KMUH). All images were acquired digitally and following local Patient Health Information (PHI) policy to extract all the private information.

To unify and perform quality control of these initial image data sets, and execute preliminary image quality screening, patients’ fundus image data were then re-exanimated to exclude over-exposure and under-exposed images which possessed no clinical values. Then the remaining qualified images were further improved by removing text information and noisy background. Pre-processing method was adopted through the adaptive histogram equalization (CLAHE) algorithm to improve the contrast of acquired images. The pre-processed images were then divided into training, validation, and test sets, respectively. In order to make statistically valuable models, we then rotated the data sets with a positive and negative 45° matrix to generate extra image information. These extra image data sets were also passed through AI models such as VGGNet, ResNet, and Transformer, as well as deep learning architectures such as ConvMixer to build predictive models as the original data sets. The RAdam (Ratified Adam) optimization tool was used as the parameter optimizer for the training of the models to accomplish a robust parameter update process. Indicators such as accuracy (ACC), area under the ROC curve (AUC) as well as Positive Predictive Value (PPV) were applied to evaluate the individual model rigidity. The research flow chart was observed and defined in Fig. 1.

The research work was coded and analyzed in python language (version 3.8.12; Python Software Foundation, Beaverton, OR USA) with provided functions such as image preprocessing and data visualization packages. Tensorflow (version 2.7.1; The Keras library in Google LLC, Mountain View, CA, USA) was also applied as a solid framework for implementing deep learning models as a necessary framework.

Within this study, two various source data sets were collected from the clinical practice information from the diabetes care generated by the Kaohsiung Datong Hospital (KMTTH) and the Endocrinology and Metabolism Clinic of the Affiliated Hospital of Kaohsiung Medical University (KMUH). The collected time span was from 2013 to 2017, respectively. During the five years’ time, for patients with Type II diabetes who had undergone clinical routine treatments and tested by routine ophthalmoscopy, the severity scores of peripheral neuropathy according to nerve conduction velocity (NCV) standard were generated and documented. Classification of the severity of peripheral neuropathy was well defined after receiving nerve conduction tests as follows: Class 0: Patients without DPN symptoms; Class 1: Patients with mild DPN symptoms; Class 2: Patients with moderate DPN symptoms; Class 3: Patients with severe DPN symptoms, retinal fundus images of the four DPN severity levels were shown in Fig. 2.

The retinal fundus images of selected clinical diabetic patients were collected by two institutes. All images were taken by the digital fundus cameras manufactured by Colin Instruments Co., Ltd. With the canon CR-2 model (Canon Medical Systems, Tustin, CA, USA). The CR-2 model has a very high resolution of up to 18 million pixels per frame. Canon Retinal Imaging Control Software (RICS) was utilized as an accompanying image control software to process and store patient fundus images with DICOM and JPEG formats for image processing needs.

Patient characteristics whose images acquired in this protocol study were shown in Table 1. A total number of patients admitted was 751, among them, 317 cases were from Datong Hospital, Kaohsiung City and 434 cases were from the Affiliated Hospital of Kaohsiung Medical University, respectively. This study was approved by the Institutional Review Board (IRB) for Affiliated Hospital of Kaohsiung Medical University. Images were collected and analyzed in accordance with relevant regulations and IRB trial guidelines set by hospitals with approval number KMUHIRB-E(I)-20190448. The informed consent is waived by the ethics committee.

Since the study number of patients with severe peripheral neuropathy was significantly lower compared to the other three categories statistically, we have decided to combine mild and severe cases into one accumulated category. This major difference certainly would create bias which introduced the inability to successfully establish AI deep learning models. Therefore, we decided that patients with severe DPN symptoms (Raw_Class 3) data shall be incorporated into patient data with moderate DPN symptoms (Raw_Class 2). So, the Class 2 definition in our review, was fine-tuned to represent patients with both moderate and severe DPN symptoms.

Table 1

Datasheet of patient characteristics
State Features	Class _Total n = 751 (100%)	Class ₀ n = 275 (37%)	Class ₁ n = 246 (33%)	Class ₂ n = 230 (30%)
State Features	Class _Total n = 751 (100%)	Class ₀ n = 275 (37%)	Class ₁ n = 246 (33%)	Raw__{Class 2} n = 176(23%)	Raw__{Class 3} n = 54(7%)
age
average value	60.70	59.05	62.00	61.79	59.70
range	21–93	21–85	27–93	32–90	31–81
gender
male	416	122	148	103	43
female	335	153	98	73	11
Note: Class _Total: All patients with Type II diabetes who have received case fundus images, Class ₀: Type II diabetics who have not had (None) peripheral neuropathy; Class ₁: Patients with Type II diabetes with mild peripheral neuropathy; Raw Class ₂: Moderate Type II diabetics with peripheral neuropathy; Raw Class ₃: Type II diabetics with severe peripheral neuropathy, Class ₂: Total patients combined with Type II diabetes mellitus with severe (Moderate and Severe, M&S) peripheral neuropathy.

In this study, age characteristics of the one-way Analysis of variance (ANOVA) algorithm were adopted, and the P value of this feature for the DPN severity category was calculated as 0.00866, and this value was less than the statistical significance threshold of 0.05 for the common two-tailed test. In addition, the chi-squared test of sex was used to analyze sex characteristics to determine the severity of DPNs. The calculated value was 6.9*, and the right tail probability covered by this value under the chi-square distribution of two degrees of freedom was also calculated as 0.05. The maximum threshold representation was 5.991, which was also calculated based on gender characterizations. All the values were less than the maximum threshold for the chi-square distribution and were all statistically meaningful.

Pre-screening was a preliminary selection process of the quality of the retinal fundus images in this study. After excluding the incomplete retina images from databases, such as the images with under-exposed and/or extra low exposure rate, and the images with over-exposed retina, the remaining 948 retinal fundus images were then screened and selected for analysis. The filtered image data sets were shown in Fig. 3. Images were screened sequentially for de-identification and background removal to preserve retinal fundus image details solely. This study was based on the CLAHE algorithm suggested by Swati. C et al. [11] with individually targeting the RGB channels to sharpen the subtle features in the image data sets. Finally, the new retinal fundus images with equally scaled superimposed three RGB channels were adopted as the model training data.

In this study, 948 retinal fundus images were enhanced by adding both ± 45-degree rotations of the original image data sets. This operation increased the original image number by three times the original, and finally, 2844 images were used in this study. The reason for his mathematical operations was to create the diversities of original images which then could solidify and strengthen our AI algorithm learning capability. Total image data sets were further divided into ratios by 7:2:1; the scaling rationale was divided into training, verification, and test sets for the establishment of AI models, where the training processes were optimized with enough but not overkilled data points.

The batch size of the model training has a certain impact on the training model implementation. According to the study by NS Keskar et al. [12], a smaller batch size possessed a higher possibility to reduce model loss and could navigate to a better search direction for a more generalized model. With this assumption, a small match size can help the AI model train the system more efficiently. Based on the above reasons and taking into account the limitations of the hardware utilized, we adjusted the match size configuration of the model to 21 to serve the purpose. Our study used a single NVIDIA GeForce RTX 3080 graphics card (NVIDIA Corp., Santa Clara, California, USA) for computing, which has 10GB of memory.

Optimization of model parameters used the Rectified Adam (RAdam) optimizer to replace the common Adam optimizer [13] which was published by L Liu et al. in 2019 for the deep learning (DL) algorithm. The DL Model Optimizer [14] in their study concluded that using a relatively lower learning rate during the warm-up period was quite effective in reducing the variation caused by using an adaptive learning rate initially. Adjusting the learning rate based on the warm-up period could stabilize the optimization gradient to achieve a much better and more stable training process.

During the model training period, in order to slow the unwanted trend of entering an overfitting dilemma, the image data generator resided in the Keras library was used as a functional formula to achieve this goal (stabilization). This study applies this specified data generator in each batch of training, importing each testing model for our data fitting and training. Because the generator can process the original image in a random manner, the overlapping situation of the trained images in each iteration will be greatly reduced, which greatly avoids the number of model learnings with the same image in multiple training iterations. This application could effectively slow down the occurrence rate of the model accidentally entering the overfitting, which fell into the dead loop with reduced efficiency in the machine learning process.

The PT-Attention model was built with the Inception v3 model [15] as a pre-training mechanism, while the “self-attention layers” were added to the original model architecture [16] as a final layer to retrain and build a robust predictive model for DPN analysis.

Our study utilized the algorithm created by A Trockman et al. who proposed the ConvMixer architecture [17] in order to establish a DPN prediction model in 2022. This predictive model’s architecture was classified as ideologically similar to ViT (Vision Transformer) [18] with a more basic MLP-Mixer (Multilayer perceptron-mixer) [19]. This ConvMixer model proposed in this study used standard convolutional layers to achieve image recognition capabilities through the Transformer architecture [16].

The research tool ResNet (Residual Network) was proposed in 2016 by Kaiming He et. al [20]. The residual architecture proposed in their study was named the “identify mapping” model [21]. This architecture alleviated the difficulty of gradient diminishing or weakening issues inside neural network structures [20].

VGG16 applied in this study consisted of 13 convolutional layers and 3 end layers connected altogether [22]. The theoretical foundation of VGG16 acted as the basis for “transfer learning” [23]. The application of VGG16 was to construct the input data with retinal fundus images of diabetic patients in our research, therefore, successfully establishing a robust classification model with the proper input data sets.

3.1. Evaluation of the predictive efficacy of the classification model of diabetic peripheral neuropathy

With modern AI algorithms, fundus photographic images of diabetic patients were used to further create a ternary classification prediction model of the severity of DPN complications. Successful and meaningful analysis was established by using the above four deep learning models. Sequential evaluation models in patients with Type II diabetes mellitus were none, mild patients with mild DPN symptoms, and moderate and severe DPN. Predicted performance of symptomatic patients (Moderate and Severe, M&S). The predictive efficacy metrics used include accuracy, sensitivity, specificity, PPV, F1 score, and AUC. The predictive parameter summary of each model on the test set is shown in Table 2.

Table 2

Evaluation summary of predictive efficacy of DPN classification models with test sets
Model	ACC	Specificity	Sensitivity	PPV	F1 score	AUC
PT-Attention	0.94	1.00	0.84	0.94	0.89	0.92
Class 1			0.73	0.89	0.80	0.86
Class 2			0.95	0.99	0.97	0.98
ConvMixer	0.90	0.92	0.90	0.80	0.83	0.93
Class 1			0.94	0.65	0.77	0.94
Class 2			0.85	0.96	0.90	0.92
ResNet34	0.97	1.00	0.90	0.98	0.94	0.95
Class 1			0.82	1.00	0.90	0.91
Class 2			0.99	0.95	0.97	0.99
PT-VGG16	0.96	0.98	0.92	0.91	0.92	0.96
Class 1			0.88	0.85	0.87	0.93
Class 2			0.96	0.98	0.97	0.98
Note: Class 1: indicates that the severity of DPN in diabetic patients - mild, and Class 2: indicates DPN in diabetic patient’s moderate and severe (M&S) categories, Sensitivity, PPV, F1 score and AUC were the average values of Class 1 and Class 2 of each model.

3.2. Confusion matrix of ternary classification model of DPN

PT-Attention, ConvMixer, ResNet-34, PT-VGG16, and other four DPN classification models for the confusion matrix of the test data set. This is shown in Fig. 4.

In the early days, due to a lack of proper knowledge of diabetic peripheral neuropathy (DPN) symptoms, the incidence of DPN increased dramatically among diabetic patients and has resulted in a loss of quality of life for diabetic cases. In fact, DPN such as foot ulcers and neuropathy could be prevented and well treated through appropriate management by improving health advocacy. To achieve this goal, our approach in this research has developed a new AI-based DPN detection mechanism; under the big data correlation process with AI model prediction, we are hoping to assist physicians in the diagnosis and screening of diabetic patients with timing efficiency for DPN, as to avoid the risk of subsequent limb ulcers associated with diabetes.

The predicting model presented based on the results of this study suggests that retinal fundus imaging in patients with Type II diabetes mellitus is heavily associated with the presence of diabetic peripheral neuropathy. The four DPN prediction models established are PT-Attention, ConvMixer, ResNet34, and ResNet34 PT-VGG16, where the accuracy of the PT-VGG16 model of the test set in this study reached a high score of 0.96; Specificity reached 0.98; The combined sensitivity of mild and moderate to severe DPN reached a level of 0.92.

Based on the association between retinal fundus imaging and diabetic peripheral neuropathy with the deep learning process, Dr. Cervera et al. [24] also showed that diabetic peripheral neuropathy (DPN) can be successfully identified by applying retinal fundus imaging alone with deep learning. The neural network Squeezenet [25] built by the authors achieved the best predictive power of all the models he studied, and finally achieved a high AUC value of 0.7097 in the test dataset. Furthermore, in the study by J Benson et. al [26], they built a binary classification model with SVMs, eventually achieving a sensitivity of 0.78 and a specificity of 0.95, respectively. These findings of Dr. Cervera and J Benson et al. also show that retinal fundus imaging in patients with diabetes can be used to predict DPN successfully.

In our study, first, we took the form of CLAHE pre-processing of retinal fundus images. Through the local processing technique, images then achieved the goal of retaining the details in the original image, while improving the overall brightness of the image itself. Compared to the original histogram equalization method, CLAHE achieved the improvement of contrast while suppressing high-frequency noises from the original image. In the meantime, this technique also ensured that there is no color banding deterioration. In this study, we limited the contrast ratio to 2.0 with the local processing matrix size of 8*8, this is our best approach in calculations, though we did not rule out that there might be other better parameter configurations that could further optimize the image quality.

The limitation of our study is that our DPN rating is completely based on the nerve conduction test (NCV) as the gold standard. Nerve conduction testing was once the standard method used to diagnose DPN, but in recent years it has been challenged as its sensitivity in detecting abnormalities in large nerve fibers but not relatively in small nerve fibers that occur in the early stages of DPN [4–6, 27]. Therefore, for AI models to produce more robust prediction results in a small never fiber obstruction, a more comprehensive diagnostic approach may be needed as a basis for classifying the severity of DPN in future correlation analysis.

In this study, an AI-assisted diagnostic model of the severity of diabetic peripheral neuropathy (DPN) was established by deep learning in the field of artificial intelligence in medicine. Through retinal fundus imaging of Type II diabetes patients, the DPN detection method, only needs to be focused on the retinal fundus image obtained by the patient. After the routine ophthalmoscopy examination, these AI image analysis tools and estimation models provide valuable information to determine whether the patient has DPN and its relative severity. This additional information contribution from the retinal fundus image also increases the physician's ability to diagnose the developmental stage of DPN as early as possible. The validity of the AI models can speed up the diagnostic process of physicians performing neurological function tests on patients, as an important factor for disease management.

Acknowledgments: This study was supported financially, in part, by grants from the Ministry of Science and Technology, Executive Yuan, Taiwan, ROC (MOST Grant numbers: 110-2221-E-992 -005 -MY2, and 111-2221-E-992 -016 -MY2).

Institutional Review Board Statement: Institutional Review Board (IRB) approval was obtained from Kaohsiung Medical University Hospital (approval number: KMUHIRB-E(I)-20190448). The IRB waived informed consent requirement of the study because it was a retrospective work. Because of legal restrictions and ethics, the data in this manuscript are available upon formal request from the corresponding author.

Informed Consent Statement: All authors have confirmed the manuscript and approved the publication of the manuscript.

Conflicts of Interest: The authors declare no conflict of interest. Part result of this study has been reported in a master thesis (Hong-Zhi Huang) in Chinese and submitted to PARINOV-PHENMA-2023 as abstract form.

Data availability: The datasets generated and analyzed during the current study are not publicly available due to privacy or ethical restrictions but are available from the corresponding author on reasonable request.

Approval statement: A statement to confirm that all experimental protocols were approved by Kaohsiung Chang Gung Medical Foundation Human Body Experiment Committee.

A. J. Boulton, "Management of diabetic peripheral neuropathy," Clinical diabetes, vol. 23, no. 1, pp. 9-15, 2005.
P. R. Vas, S. Sharma, and G. Rayman, "Distal sensorimotor neuropathy: improvements in diagnosis," The review of diabetic studies: RDS, vol. 12, no. 1-2, p. 29, 2015.
S. Tesfaye et al., "Diabetic neuropathies: update on definitions, diagnostic criteria, estimation of severity, and treatments," Diabetes care, vol. 33, no. 10, pp. 2285-2293, 2010.
A. I. Vinik, M.-L. Nevoret, C. Casellini, and H. Parson, "Diabetic neuropathy," Endocrinology and Metabolism Clinics, vol. 42, no. 4, pp. 747-787, 2013.
A. Breiner, L. E. Lovblom, B. A. Perkins, and V. Bril, "Does the prevailing hypothesis that small-fiber dysfunction precedes large-fiber dysfunction apply to type 1 diabetic patients?," Diabetes care, vol. 37, no. 5, pp. 1418-1424, 2014.
J. C. Won and T. S. Park, "Recent advances in diagnostic strategies for diabetic peripheral neuropathy," Endocrinology and metabolism, vol. 31, no. 2, pp. 230-238, 2016.
Z. Li et al., "Deep learning methods for lung cancer segmentation in whole-slide histopathology images—the acdc@ lunghp challenge 2019," Journal of Biomedical and Health Informatics (JBHI), vol. 25, no. 2, pp. 429-440, 2020.
G. Wang et al., "A deep-learning pipeline for the diagnosis and discrimination of viral, non-viral and COVID-19 pneumonia from chest X-ray images," Nature biomedical engineering, vol. 5, no. 6, pp. 509-521, 2021.
M. H. Megallaa, A. A. Ismail, M. H. Zeitoun, and M. S. Khalifa, "Association of diabetic foot ulcers with chronic vascular diabetic complications in patients with type 2 diabetes," Diabetes & Metabolic Syndrome: Clinical Research & Reviews, vol. 13, no. 2, pp. 1287-1292, 2019.
A. Shahidi et al., "Retinal nerve fibre layer thinning associated with diabetic peripheral neuropathy," Diabetic Medicine, vol. 29, no. 7, pp. e106-e111, 2012.
C. Swathi, B. Anoop, D. A. S. Dhas, and S. P. Sanker, "Comparison of different image preprocessing methods used for retinal fundus images," in Conference on Emerging Devices and Smart Systems (ICEDSS), 2017: IEEE.
N. S. Keskar, D. Mudigere, J. Nocedal, M. Smelyanskiy, and P. T. P. Tang, "On large-batch training for deep learning: Generalization gap and sharp minima," in International Conference on Learning Representations (ICLR), 2017.
D. P. Kingma and J. Ba, "Adam: A method for stochastic optimization," in International Conference on Learning Representations (ICLR), 2015.
L. Liu et al., "On the variance of the adaptive learning rate and beyond," in International Conference on Learning Representations (ICLR), 2020.
C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, "Rethinking the inception architecture for computer vision," in Conference on Computer Vision and Pattern Recognition (CVPR), 2016: IEEE, pp. 2818-2826.
A. Vaswani et al., "Attention is all you need," Advances in neural information processing systems, vol. 30, pp. 5998--6008, 2017.
A. Trockman and J. Z. Kolter, "Patches Are All You Need?," in International Conference on Learning Representations (ICLR), 2022.
A. Dosovitskiy et al., "An image is worth 16x16 words: Transformers for image recognition at scale," in International Conference on Learning Representations (ICLR), 2021.
I. O. Tolstikhin et al., "Mlp-mixer: An all-mlp architecture for vision," Advances in Neural Information Processing Systems, vol. 34, pp. 24261-24272, 2021.
K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Conference on Computer Vision and Pattern Recognition (CVPR), 2016: IEEE, pp. 770-778.
K. He, X. Zhang, S. Ren, and J. Sun, "Identity mappings in deep residual networks," in European Conference on Computer Vision (ECCV), 2016: Springer, pp. 630-645.
K. Simonyan and A. Zisserman, "Very deep convolutional networks for large-scale image recognition," in International Conference on Learning Representations (ICLR), 2015.
K. Weiss, T. M. Khoshgoftaar, and D. Wang, "A survey of transfer learning," Journal of Big data, vol. 3, no. 1, pp. 1-40, 2016.
D. R. Cervera, L. Smith, L. Diaz-Santana, M. Kumar, R. Raman, and S. Sivaprasad, "Identifying Peripheral Neuropathy in Colour Fundus Photographs Based on Deep Learning," Diagnostics, vol. 11, no. 11, p. 1943, 2021.
F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally, and K. Keutzer, "SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size," in International Conference on Learning Representations (ICLR), 2017.
J. Benson, T. Estrada, M. Burge, and P. Soliz, "Diabetic peripheral neuropathy risk assessment using digital fundus photographs and machine learning," in Engineering in Medicine and Biology Society (EMBC), 2020: IEEE.
S. Divisova et al., "Prediabetes/early diabetes‐associated neuropathy predominantly involves sensory small fibres," Journal of the Peripheral Nervous System, vol. 17, no. 3, pp. 341-350, 2012.

No competing interests reported.

Using Deep Learning Models to Predict the Risk of Peripheral Neuropathy on Diabetic Patients

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials and Methods

3. Results

3.1. Evaluation of the predictive efficacy of the classification model of diabetic peripheral neuropathy

3.2. Confusion matrix of ternary classification model of DPN

4. Discussion

5. Conclusions

Declarations

References

Additional Declarations

Status:

Version 1