From Evaluation to Prediction: Analysis of Diabetic Autonomic Neuropathy Using Sudoscan and Artificial Intelligence

Abstract

A dangerous side effect of diabetes that can significantly lower quality of life and raise the death rate of diabetic individuals is diabetic autonomic neuropathy. It is essential to identify and anticipate this disease early on for prompt intervention and care. This study aims to predict this diabetic complication using Sudoscan and artificial intelligence. In this study, 172 individuals with type 1 or type 2 diabetes mellitus provided clinical and demographic information. Sudoscan was used to evaluate the subjects’ sudomotor dysfunction. Statistical methods were used to link various electrochemical skin conductance values with risk factors for neuropathy such as age, BMI, age of diabetes, or biochemical values such as cholesterol and triglycerides. Different machine-learning algorithms were used to predict the risk of diabetic autonomic neuropathy based on the collected data. The accuracy achieved with Logistic Regression is 92.6%, and with the Random Forest model is 96.3%. Lazzy Classifiers also show that six classifiers have a high performance of 97%. Thus, the use of machine learning algorithms in this field of metabolic diseases offers new perceptions for diagnosis, treatment, and prevention, and improves the quality of life of diabetic patients by reducing the incidence of complications related to this disease.

1. Introduction

Worldwide, diabetic neuropathy is the most common complication of diabetes and the most common cause of hospitalization and non-traumatic amputations among diabetic complications. Diabetic neuropathy can occur at any time in its course and affects between 50 and 75% of patients with diabetes [1]. Diabetic neuropathy (DN) is defined as the totality of anatomical, clinical, neurological, and metabolic disorders occurring during the course of diabetes mellitus, after the exclusion of other neurological manifestations belonging to diabetes mellitus-associated or other causes (hereditary, traumatic, neoplastic, infectious, compressive, metabolic, autoimmune) or secondary to other systemic conditions. Its frequency increases with disease duration, patient age, abdominal circumference, ethanol, and neurotoxic consumption. After 20 years of progression, two-thirds of people with diabetes will have at least one clinical or subclinical manifestation of neuropathy. Distal symmetric polyneuropathy and diabetic autonomic neuropathy, especially cardiovascular neuropathy, are the most prevalent and studied clinical forms [2].

In recent years, different non-invasive techniques have been implemented for the early detection and treatment of DN. Among these, infrared thermography has shown promise as a diagnostic technique for this disease. Recent studies [3,4,5] have shown that infrared thermal imaging can identify early indicators of neuropathy by studying heat patterns in the skin.These techniques also provide important new insights into vascular and neuronal alterations related to diabetic neuropathy. However, in order to provide a comprehensive assessment of neuropathic problems in diabetic patients, our study is based on another non-invasive technique, which is Sudoscan.

The manifestation of autonomic neuropathy with peripheral consequences is sudomotor neuropathy; thus, sweat glands are sympathetically innervated and the touching of postganglionic fibers is manifested by reduced secretion of these glands, resulting in anhidrosis and dry, cracked skin, with trophic disorders in advanced stages, but also disorders of adaptation to temperature changes (frequently heat intolerance) [6,7]. Reduced/absent sweating correlates with other signs and symptoms of autonomic neuropathy, especially orthostatic hypotension, but also with other cardiovascular changes (loss of heart rate variability). The association between the existence of non-physiological pressure points and hyposudoration, anhidrosis, and dry skin will lead, over time, through repeated mechanical overstrain, to the appearance of calluses, cracks, or other areas of continuity which are real entry points for various pathogens [8].

Assessment of peripheral autonomic neuropathy can be performed by the sympathetic skin reaction: an examination of sympathetic efferent using sweating produced by tension changes (“galvanic skin reflex”). This analysis can be performed with Sudoscan, equipment also used in this study to assess sudomotor dysfunction by measuring skin conductance, providing valuable information on DAN risk. This medical device provides a non-invasive assessment of sweat gland function on the palms and soles of the feet, areas with the highest density of sweat glands. The reliability and validity of Sudoscan have been demonstrated in several articles [9], e.g., Carolina M Casellini demonstrated the correlation between electrochemical skin conductance (ESC) and clinical, somatic, and autonomic measures of neuropathy [10], Smith et al. showed that it is a test that promises to diagnose DAN with similar intraepidermal nerve fiber density performances, but cannot replace the diagnosis made with electromyography [11]. Selvarajah et al. emphasized the importance of examining patients with Sudoscan to improve individualized management of diabetics through early detection and intervention [12]. Thus, early recognition of DAN symptoms is essential for the management of diabetes and related complications.

The hypothesis of this study is based on the fact that the use of Sudoscan equipment together with artificial intelligence algorithms can improve DAN diagnosis and prediction.

Approaching this condition through the lens of Sudoscan and artificial intelligence is very poorly addressed in the literature. Solatidehkordi Z. and Dhou S. employed different machine learning models, like support vector machines, Logistic Regression, Random Forest, Extreme Gradient Boosting, and CatBoost, in order to identify DAN in diabetic patients by analyzing relevant data from electronic health records, including Sudoscan data [13].

Instead, artificial intelligence methods have been applied and highlighted in numerous studies for the diagnosis and management of diabetes [14,15]. Several papers have focused on demonstrating the importance of artificial intelligence in predicting various diabetes complications [16,17] and a few have also focused on DAN [18,19,20] by analyzing corneal confocal microscopy.

Our study has made the following contributions: (i) propose an early prediction of DAN for diabetic patients, (ii) use a private dataset that was gathered from Suceava/Romania citizens, (iii) analyze the effectiveness of several machine learning models in detecting neuropathy using Sudoscan and fundamental clinical data, (iv) obtain high accuracy, sensitivity, specificity, and AUC by the predictive models created in this work, demonstrating strong performance.

Thus, new artificial intelligence technologies together with Sudoscan may represent a promising potential in research and prevention of diabetes complications, offering new horizons for improving the quality of life of patients with diabetes.

2. Materials and Methods

2.1. Participating Subjects

This study involved 172 patients with diabetes mellitus (type 1 and 2) who presented to the Diabetes, Nutrition and Metabolic Diseases Department and outpatient department of the “Sfântul Ioan cel Nou” Clinical Hospital in Suceava. The protocol of this study was approved by the Ethics Committee of the Hospital. The inclusion criteria consisted of patients older than 18 years old with a diagnosis of type 1 or 2 diabetes who provided informed consent. The subjects excluded from this study are those with skin conditions that affect the hands or the feet, patients with electrical implants (pacemakers, insulin pumps), and those without the capacity to give informed consent. After signing the informed consent documents, the patients underwent non-invasive Sudoscan testing. The Sudoscan test was performed according to the manufacturer’s protocol by qualified medical personnel. Patients placed their hands and feet on the stainless-steel electrodes of the device. Sudoscan equipment provides information on electrochemical skin conductance (ESC). These measurements provide valuable information on the sudomotor dysfunction that is associated with diabetic autonomic neuropathy. The testing procedure is non-invasive, safe, takes about 2–3 min, and is performed only once for each patient. The SUDOSCAN device produces discrete measurement values. A session with the SUDOSCAN device provides instantaneous ESC values for the hands and feet.

We also collected demographic and clinical data such as age, height, weight, year of diabetes diagnosis, cholesterol, triglycerides, blood pressure, creatinine, and glycosylated hemoglobin (HbA1c) from the patients. The data collected have been anonymized and stored securely, respecting data protection regulations and participants’ privacy.

2.2. Sudoscan Device

Sudoscan [21] is a non-invasive device that accurately assesses sweat gland function in the palms of the hands and soles of the feet, areas with the highest density of sweat glands. Sweat glands are innervated by small sympathetic C-fibers. Sweat dysfunction is one of the earliest neurophysiological signs detectable in distal small-fiber neuropathy, and quantitative assessment of the sweat response is an indicator of the severity and distribution of autonomic dysfunction.

Very low voltage levels (<4 V) stimulate sweat glands, producing a measurable flow of ionic activity through sweat ducts. Electrochemical reactions between electrodes and measurements of the amount of chlorine in the sweat glands are recorded and analyzed.

This method allows a highly reproducible quantitative assessment of sweating disorders, which is not possible under normal physiological conditions. Quantitative results assessing skin conductance (ESC) are expressed in microSiemens (μS) for hands and feet. In addition, SUDOSCAN provides an overall risk score derived from local ESC values and demographic data.

Sudoscan provides information on ESC for hands and feet. Thus, ESC values > 60 μS indicate patients without diabetic neuropathy, ESC values between 40 and 60 μS indicate patients with possible diabetic neuropathy for the hands, and between 40 and 50 for the feet; ESC values < 40 for the hands and ESC < 50 for the feet indicate advanced diabetic neuropathy. DAN can lead over time to diabetic nephropathy and impaired cardiovascular function. So, in addition to a neuropathy risk score, this equipment also provides us with nephropathy risk and cardiac risk. Typical ranges for risk scores provided by Sudoscan are as follows: a score of 50 or higher is usually considered normal, scores between 40 and 50 may suggest mild sweat gland dysfunction and scores below 40 could indicate significant sweat gland dysfunction.

2.3. Statistical Analysis

Various statistical methods were used to sample the data collected from diabetic patients. Descriptive statistics including calculation of mean, standard deviation, and minimum and maximum values were applied to the whole database. Percentages provide a more detailed view of the distribution of the data. Different types of graphs are used to easily visualize the distribution of ESC and identify outliers, depending on the neuropathy status. Spearman correlation analysis is applied to determine the relationship between ESC (both hand and foot) and other demographic or clinical data collected. This method of correlation was preferred over Person correlation because it does not require normally distributed data, and we found that it is preferred by most researchers who have studied ESC using Sudoscan [22,23,24]. Through these statistical analyses, this study provides a proper view of the collected data and the distribution of ESC according to DAN status and various clinical and demographic variables.

2.4. Machine Learning Algorithms for Prediction

Machine learning models play an important role in predicting various diseases, including diabetes and its complications. These methods also contribute to the early identification of patients at high risk of DAN and allow early interventions for disease management and improvement of the patient’s quality of life. DAN is influenced by a number of factors such as duration of diabetes, patient age, lifestyle, and glycaemic and lipid control. Machine learning algorithms can analyze very large sets of clinical and/or biological data to identify relationships between different factors and the development of DAN. With these methods, predictive models can even be developed to estimate the likelihood of a patient developing DAN based on clinical or laboratory data alone.

Thus, in this study, different algorithms were used such as Logistic Regression (LR) [25], a linear model for classification selected for its efficacy in binary classification tasks and its capacity to yield probabilistic results, and Random Forest (RF) [26], which uses multiple decision trees to improve classification results selected because of its resilience to overfitting, robustness, and capacity to handle non-linear interactions. The Lazy classifier [27] model from the lazypredict library [28] was also applied, a very useful tool to simplify the process of comparing different types of algorithms. This method is one that provides a quick view of multiple types of classifiers without requiring detailed configuration of each. In this study, the Lazzy method tests several models:

• Linear SVC (Linear Support Vector Classification) [29]—a classifier of SVM type which uses linear kernel for classifying tasks.
• Linear Discriminant Analysis (LDA) [30]—derives class conditional densities according to Bayes’ rule and gives linear decision boundaries.
• Calibrated Classifier CV (corss-validation) [31]—applies cross-validation to calibrate the base classifier in order to produce probabilistic predictions.
• Ridge Classifier CV (corss-validation) [32]—classification model with built-in cross-validation to find the best value of regularization parameter.
• Ridge Classifier [28]—ridge regression model tailored for classification.
• Passive Aggressive Classifier [33]—an online learning algorithm that is significantly fast and can deal with large-scale training problems.
• SGD Classifier (Stochastic Gradient Descent Classifier) [34]—stochastic gradient descent-based linear classifier.
• Perceptron [35]—online learning algorithm based on an artificial neural network.
• Logic Regression (Logistic Regression) [25]—statistical model that uses logistic function handling binary dependent variable.
• LGBM Classifier (Light Gradient Boosting Machine) [36]—a learning technique that uses tree-based learning algorithms.
• Extra Trees Classifier (Extremely Randomized Trees Classifier) [37]—a learning method that combines the outcomes of numerous decision trees, randomly drawn.
• Bernoulli NB (Bernoulli Naive Bayes) [38]—Naive Bayes classifier for binary/boolean features.
• Decision Tree Classifier [39]—classifier that uses a tree-like model of decisions and their possible consequences.
• Nearest Centroid [40]—classifier that assigns to each sample the label of the closest centroid.
• Extra Tree Classifier [41]—single decision tree that is part of the Extremely Randomized Trees ensemble.
• XGB Classifier (Extreme Gradient Boosting) [41]—a fast and efficient gradient-boosted decision tree implementation.
• Random Forest Classifier [26]—a learning technique that uses averaging to increase predictive accuracy after fitting several decision tree classifiers on different subsamples.
• Ada Boost Classifier [42]—a learning method that combines multiple weak classifiers to create a strong one.
• Bagging Classifier [43]—a learning method that fits several model iterations on random subsamples of the dataset and then averages the predictions.
• Nu SVC (Nu Support Vector Classification) [44]—a variant of SVM that uses a parameter nu to control the number of support vectors.
• SVC (Support Vector Classification) [45]—another variant of SVM that employs linear or non-linear classification based on various kernel functions.
• Gaussian NB (Gaussian Naive Bayes) [38]—a probabilistic ML algorithm used for many classification functions, based on the Bayes theorem.
• K-Neighbors Classifier (K-Nearest Neighbors Classifier) [46]—a non-parametric method that uses the nearest neighbors’ majority vote for classification purposes.
• Quadratic Discriminant Analysis (QDA) [47]—a classifier similar to LDA which permits each class to have a covariance matrix.

3. Results

3.1. Statistical Analysis

Table 1. Characteristics of the data collected from the 172 subjects.

	Mean (std)	25th Percentile	75th Percentile
Age	60.9 (11.47)	55.0	69.0
BMI	31.73 (6.15)	27.5	35.0
Left foot	62.8 (15.36)	50.0	75.0
Right foot	61.4 (15.78)	49.0	74.0
Left hand	57.5 (14.73)	48.0	69.0
Right hand	57.8 (14.38)	49.0	69.0
Neuro risk	60.0 (13.19)	50.2	71.3
Cardio risk	37.3 (11.37)	30.0	43.0
Nefro risk	61.5 (18.65)	49.0	71.0
DM age	11.02 (8.63)	4.0	16.0
HbA1C	8.73 (2.00)	7.2	9.9
Cholesterol	202.9 (48.76)	172.7	233.3
Triglyceride	159.0 (81.2)	103.0	185.0
SBP	143.0 (20.13)	129.0	156.3
DBP	80.0 (12.06)	70.0	87.3
Creatinine	0.9 (0.25)	0.6	1.0
Percentage of Type 1 Diabetes	12.21%

BMI—Body Mass Index; DM age—the age of Diabetes Mellittus; HbA1C—glycated hemoglobin; SBP—Systolic Blood Pressure; DBP—Diastolic Blood Pressure.

shows the mean values and standard deviations and 25th and 75th percentiles for all variables collected in the database. This result was obtained using pandas and numpy library in Phyton 3.9.

Analyzing the statistical results, we can describe the overall database used in this study. The mean age of the subjects is 60.9 (11.47 std) and they have a mean BMI of 31.73 (6.15 std), indicating an obesity-prone group. The mean ESC values for hands and feet place the group in the one without the risk of neuropathy. However, it should be taken into account that patients with neurotrophic treatment (21% of all subjects) were included in these data.

The mean age at diabetes diagnosis is 11.02 years, suggesting that patients were diagnosed with diabetes around the age of 50. Glycated hemoglobin (HbA1c) is an indicator of long-term glycemic control and has a mean of 8.73%, indicating poor glycemic control in this group of subjects. The mean values for SBP and DBP show a tendency of this group towards hypertension. The mean values for cholesterol and triglyceride classify the group as hypercholesterolemic and hypertriglyceridemic, and the standard deviation values show the high variability of these biochemical values among the target group.

In Figure 1 the boxplots represent the Electrochemical Skin Conductance (ESC) distribution for left hand/leg and right hand/leg, grouped according to diabetic autonomic neuropathy (DAN) status: possible DAN, confirmed DAN, and no DAN.

In

Table 2. Interpretation of BMI for the ESC results obtained for feet/hands.

		ESC Feet			ESC Hands
	Weight Status	NAD	No NAD	PossibleNAD	NAD	No NAD	Possible NAD
No tratment	Normal weight	2	16	2	1	15	4
	Overweight	7	21	5	4	17	12
	Obese	30	29	23	15	25	42
Neurotrophic tratment	Normal weight	1	1	0	1	1	0
	Overweight	1	10	1	2	6	4
	Obese	4	16	3	1	13	9

, an analysis of the ESC results obtained by patients in both hands and feet according to BMI and separated for those on and off neurotrophic treatment was performed. The Interpretation column represents the BMI coding as follows: 0—normal weight, 1—overweight, and 2—obese. Interpretation of ESC results is as follows: ESC values > 60 μS indicate DAN, ESC values between 40 and 60 μS indicate possible DAN for hands and between 40 and 50 for feet; ESC values < 40 for hands and ESC < 50 for feet indicate NO DAN. Thus, it can be seen that there is a higher incidence of DAN or possible DAN for obese or overweight patients, which is in agreement with the literature [22] which states that a high weight status leads to early complications in patients with diabetes. This tendency to overweight and obesity is present in 61% of all patients in this study. Neurotrophic treatment clearly improves SUDOSCAN values.

Table 3. Interpretation of cholesterol values for the ESC results obtained for feet/hands.

		ESC Feet			ESC Hands
	CHO * Values	NAD	No NAD	PossibleNAD	NAD	No NAD	Possible NAD
No tratment	<180 mg/dl	1	31	5	5	23	9
	(180, 200) mg/dl	4	9	3	1	9	6
	>200 mg/dl	32	26	24	31	25	26
Neurotrophic tratment	<180 mg/dl	0	14	0	2	9	3
	(180, 200) mg/dl	1	3	0	1	3	0
	>200 mg/dl	5	10	4	6	8	5

* cholesterol.

represents an analysis of the ESC results obtained by patients in both hands and feet according to neurotrophic treatment (yes—1 or no—0) and total cholesterol values. Cholesterol values were divided into three groups: 0—values < 180 mg/dl, 1—values between 180 and 200 mg/dl, and 2—values greater than 200 mg/dl. In the no-treatment group, most patients with DAN are in category 2 (total cholesterol values: 200 mg/dl).

Table 4. Interpretation of values for the ESC results obtained for feet/hands.

		ESC Feet			ESC Hands
	TG * Values	NAD	No NAD	Possible NAD	NAD	No NAD	Possible NAD
No tratment	<180 mg/dl	8	34	11	7	31	15
	(180, 200) mg/dl	1	6	5	1	3	8
	>200 mg/dl	26	26	16	29	23	18
Neurotrophic tratment	<180 mg/dl	1	11	0	1	8	3
	(180, 200) mg/dl	1	4	0	0	5	0
	>200 mg/dl	4	12	4	8	7	5

* triglyceride.

represents an analysis of the ESC results obtained by patients in both hands and feet according to neurotrophic treatment (yes—1 or no—0) and triglyceride values which were divided into three groups: 0—values < 130 mg/dl, 1—values between 130 and 150 mg/dl, and 2—values greater than 150 mg/dl. In the no-treatment group, it is observed that most patients with DAN are in category 2, similar to the previous situation. As suggested by Tesfaye [48], in addition to improving glycemic control, the identification of potentially modifiable risk factors for diabetic neuropathy is of particular importance. EURODIAB identified statistically significant positive correlations between the prevalence of diabetic neuropathy and duration of diabetes, triglyceride levels, presence of hypertension, and age. In our study group, hypercholesterolemia and hypertriglyceridemia were associated with the occurrence of the condition, indicating the prevalence of mixed dyslipidemia in diabetic patients. Thus, dyslipidemia, through the factors followed (triglyceridemia and cholesterolemia), is a risk factor for the development of diabetic neuropathy.

Figure 2 shows the distribution of ESC values according to the age of diabetes diagnosis divided into three categories: 0–5 years, 6–10 years, and more than 10 years after diagnosis. In the case of patients included in the DAN or possibly DAN categories, according to the ESC values obtained with SUDOSCAN for hands and feet, it is clear that most patients have more than 10 years since they were diagnosed with this metabolic disease.

In

Table 5. Spearman correlations and significance level.

	Hands			Feet
	r	p	95% CI	r	p	95% CI
Age	−0.14	0.073	[−0.65, −0.23]	−0.09	0.224	[−0.66, −0.24]
BMI	−0.20	0.009	[−0.35, −0.05]	−0.31	<0.001	[−0.45, −0.17]
Neuro risk	−0.89	<0.001	[−0.95, −0.83]	−0.89	<0.001	[−0.95, −0.83]
Cardio risk	−0.08	0.305	[−0.23, 0.07]	−0.23	0.003	[−0.37, −0.09]
Nefro risk	−0.24	0.001	[−0.38, −0.10]	−0.30	<0.001	[−0.44, −0.16]
DM age	−0.23	0.003	[−0.50, −0.16]	−0.15	0.047	[−0.52, −0.18]
HbA1C	0.04	0.613	[−0.13, 0.21]	−0.02	0.809	[−0.19, 0.15]
Cholesterol	−0.31	<0.001	[−0.15, 0.19]	−0.48	<0.001	[−0.46, 0.16]
Triglyceride	−0.30	0.001	[−0.34, −0.04]	−0.26	0.001	[−0.37, −0.07]
SBP	−0.04	0.573	[−0.21, 0.13]	−0.17	0.028	[−0.32, −0.02]
DBP	−0.01	0.947	[−0.18, 0.16]	−0.04	0.588	[−0.21, 0.13]
Creatinine	−0.02	0.747	[−0.19, 0.15]	0.04	0.568	[−0.13, 0.21]

BMI—Body Mass Index; DM age—Age of Diabetes Mellitus diagnosis; HbA1C—glycated hemoglobin; SBP—Systolic Blood Pressure; DBP—Diastolic Blood Pressure.

, we can identify notable correlations between the variables examined and susceptibility to diabetic neuropathy by interpreting the Spearman correlation values and significance level for ESC assessed for both hands and feet in the context of diabetic neuropathy risk. Thus, statistically significant and negative correlations are observed between ESC values for both hands and feet and values for neurological risk. The correlation coefficient −0.89 indicates a strong negative correlation and that, as ESC decreases, the risk of diabetic neuropathy increases considerably. Patient age and age of diabetes are also negatively correlated with ESC values for hands and feet. The coefficient of correlations between the age of the patients and values of ESC for Hands and Feet (−0.14 and −0.09, respectively) and Age of Diabetes Mellitus diagnosis (DM age), with ESC values of −0.23 and −0.15, respectively, indicate week negative correlations, suggesting that age is inversely proportional to ESC, i.e., older people with older diabetes have lower ESC scores, suggesting a higher predisposition to neuropathy.

Body Mass Index (BMI) correlated with Hands and Feet values for ESC (−0.20 and −0.31, respectively), suggest that higher BMI is associated with lower ESC scores, indicating a possible increased predisposition to diabetic neuropathy in individuals with higher body weight. Also, moderately significant negative correlations indicate that an increase in nephrological risk and triglyceride levels is associated with lower ESC scores (Nephro risk and Triglyceride with Hands_r and Feet_r: −0.24 and −0.30; −0.30 and −0.26).

The correlation data assist in determining which variables have the strongest associations with DAN. This helps us to understand which features are probably crucial for developing machine learning models.

3.2. Prediction Using Machine Learning Methods

Codes were implemented in Phyton 3.9 using two different methods to classify the data: LR and RF. These methods are applied to the dataset collected from patients without neurotrophic treatment (135 subjects) (‘Age’, ‘BMI’, ‘Avg feet’, ‘Avg hands’, ‘DM age’, ‘HbA1c’, ‘Cholesterol’, ‘Triglyceride’, ‘SBP’, ‘DBP’, ‘Creatinine’) to predict the risk of diabetic autonomic neuropathy based on the target variable represented by the mean values obtained for hands and transformed into binary. This decision, to exclude patients with neurotrophic treatment, was made to ensure that the models were trained on data reflecting the risks of diabetic complications without the influence of treatment effects.

We additionally employed the Lazzy Predict library to create various machine-learning methods. This machine-learning Python package allows us to rapidly determine which models, without manually configuring and testing each one, could work best for our dataset. Hundreds of models can be tested to find which one fits our dataset the best. With its display of multiple metrics like accuracy, F1 score, and ROC AUC, Lazzy Predict provides a clear picture of the performance of several models.

Explanation of the stages of the used code:

• The first classification model used is a linear one: Logistic Regression. Then was applied Random Forest, a model that uses multiple decision trees to improve classification performance. To evaluate the performance of the classifiers we used the following: confusion matrix, roc_auc_score, and accuracy_score. The confusion matrix gives a complete overview of the classifier performance by comparing predictions with actual values, and the accuracy_score quantifies the percentage of correct predictions. Roc_auc_score assesses the model’s ability to correctly classify positive and negative predictions at different thresholds.
• Preparation of data for classification: all collected values (‘Age’, ‘BMI’, ‘Avg feet’, ‘Avg hands’, ‘DM age’, ‘HbA1c’, ‘Cholesterol’, ‘Triglyceride’, ‘SBP’, ‘DBP’, ‘Creatinine’) are assigned X, and the column ‘Hands’ (average for ESC values measured for the left and right hand—binary: 0—representing the class at risk or possible risk of DAN (78 samples), 1—representing the class at no risk of DAN (57 samples)), the variable Y.
• Data are split into the training set (80%) and the test set (20%), using the train_test_split function.
• The models are initialized with the specified parameters using the two classifiers: LR and Random Forest.
• The model is trained on the training dataset using fit().
• Then, prediction is performed on the test dataset using predict().
• In the last step, the models are evaluated.

Table 6. Results obtained with Logistic Regression and Random Forest.

Algorithm	Accuracy	ROC AUC	F1 Score	Sensitivity	Specificity	Confusion Matrix
Logistic Regression	92.59%	92.50%	92.31%	91.67%	93.33%	[[14, 1], [1, 11]]
Random Forest	96.30%	95.83%	96.00%	91.67%	100%	[[15, 0], [1, 11]]

ROC AUC—Receiver Operating Characteristic Area Under the Curve.

presents the performance metrics obtained with the LR and RF classifiers. We can see that RF outperformed the LR classifier with an accuracy of 96.30%, compared to 92.59%. These findings imply that the models are trustworthy in this particular DAN prediction. The RF model yielded a ROC AUC score of 95.83%, whereas the LR model yielded a score of 92.50%. These results show that both models can accurately differentiate between positive and negative cases of diabetic autonomic neuropathy. With the RF model, we obtained an F1 score of 96.00%, whereas the LR model has a score of 92.31%. This suggests that both models can accurately predict both classes, but the RF model performs better once again.

The sensitivity for both models was 91.67%. This indicates that 91.67% of individuals at risk were correctly identified by both models as positive cases of DAN. The RF model achieved a flawless specificity of 100%, while the LR model’s specificity was 93.33%. Specificity shows how well the classifier recognizes negative cases. The RF perfect score shows that it accurately detected every negative occurrence. The confusion matrices demonstrate that both models generate relatively few errors; however, the RF model performed better because it produced no false positives.

Upon first evaluating our models, we found that RF produced 96% accuracy. Nonetheless, there may be bias introduced when evaluating the model with a single test fraction of 20%. In particular, the claimed accuracy may be too optimistic if the test set included softer cases. On the other hand, accuracy can be overestimated if the test set is extremely challenging. This fluctuation emphasizes the danger that comes with depending just on one random split of the data set.

For the two classification models, we used k-fold cross-validation in order to solve this issue and obtain a more trustworthy model performance metric. The dataset is split up into 10 subsets (folds) for k-fold cross-validation. To make sure that every data point is used for both training and testing, each model is trained on nine folds and tested on the remaining fold. This process is repeated 10 times. By using this method, the likelihood of bias is decreased, and a more precise and broadly applicable estimate of the model’s performance is obtained. K-fold cross-validation makes sure that the reported performance metrics, including accuracy and ROC AUC, reflect the model’s capacity to generalize to new data rather than relying on the features of a particular subset by employing the complete data set for evaluation in a methodical manner.

The results of the K-fold cross-validation, reported in

Table 7. K-fold cross-validation results.

Algorithm	Mean Accuracy	Mean ROC AUC
Logistic Regression	95.6%	99.75%
Random Forest	99.23%	100%

, show an average accuracy of 95.6% and an average ROC AUC of 99.75% obtained with RL, indicating strong performance and excellent class discrimination, while, with RF, we obtained an average accuracy of 99.23% and a perfect average ROC AUC of 100%, demonstrating extremely high accuracy and perfect class separation. The results show that both models are highly efficient, with RF slightly outperforming RL.

Using the same training and testing dataset created initially, LazyClassifier is applied. The first step of the Python code is the initialization of the classifier with the default parameters, then the classifier trains multiple models on the training data and makes predictions on the testing data. The performance of the models is displayed together with balanced accuracy, ROU AU, and F1 score, as can be observed in

Table 8. Results obtained with Lazzy Classifier.

Model	Accuracy	Balanced Accuracy	ROC AUC	F1 Score
Linear SVC	0.96	0.97	0.97	0.96
Linear Discriminat Analysis	0.96	0.97	0.97	0.96
Calibrated Classifier CV	0.96	0.97	0.97	0.96
Ridge Classifier CV	0.96	0.97	0.97	0.96
Ridge Classifier	0.96	0.97	0.97	0.96
Passive Aggressive Classifier	0.96	0.97	0.97	0.96
SGD Classifier	0.93	0.95	0.95	0.93
Perceptron	0.93	0.95	0.95	0.93
Logic Regression	0.93	0.95	0.95	0.93
LGBM Classifier	0.96	0.94	0.94	0.96
Extra Trees Classifier	0.96	0.94	0.94	0.96
Bernoulli NB	0.89	0.92	0.92	0.89
Decision Tree Classifier	0.89	0.92	0.92	0.89
Nearest Centroid	0.89	0.91	0.91	0.89
Extra Tree Classifier	0.93	0.91	0.91	0.93
XGB Classifier	0.93	0.91	0.91	0.93
Random Forest Classifier	0.93	0.91	0.91	0.93
Ada Boost Classifier	0.93	0.91	0.91	0.93
Bagging Classifier	0.93	0.91	0.91	0.93
Nu SVC	0.93	0.88	0.88	0.92
SVC	0.93	0.88	0.88	0.92
Gaussian NB	0.81	0.87	0.87	0.82
K-Neighbors Classifier	0.89	0.85	0.85	0.89
Quadratic Discriminant Analysis	0.78	0.77	0.77	0.78

ROC AUC—Receiver Operating Characteristic Area Under the Curve.

.

The LazzyClassifier method was also tested to see if there are other, better classifiers.

It can be seen from Table 8 that very good performances were obtained with Linear SVC, Linear Discriminate Analysis, Calibrated Classifier CV, Ridge Classifier CV, RidgeClassifier, Passive Aggressive Classifier, LGBM Classifier, and Extra Trees Classifier. Using these models, we obtained an accuracy of 96%, an F1 score of 0.96, and an ROC AUC value of 0.97 (except LGBM Classifier and Extra Trees Classifier, which had an ROC AUC of 0.94). These models have a very high ability to distinguish between positive and negative cases of diabetic autonomic neuropathy. Balanced accuracy (0.97) shows that these models perform well even on unbalanced datasets and the values obtained from the Sudoscan tests together with the other clinical and demographic data collected can be used to predict the diagnosis of diabetic autonomic neurotrophy with very good accuracy.

A performance of 93% was obtained with the classifiers SGD Classifier, Perceptron, Logic Regression, Extra Tree Classifier, XGB Classifier, RandomForest Classifier, Ada Boost Classifier, Bagging Classifier, Nu SVC, and SVC. This accuracy of 93%, is a decent one, with a value for balanced accuracy and ROC AUC of 0.91–0.95 indicating acceptable performance. The SVC and Nu SVC models had values of 0.88 for balanced accuracy and ROC AUC, suggesting a somewhat poorer performance in handling imbalances in the data.

Moderate performance (89–81% accuracy) was obtained with the classifiers Bernoulli NB, Decision Tree Classifier, Nearest Centroid, Gaussian NB, and K-Neighbors Classifier. The ROC AUC and balanced accuracy of these models ranged between 81% and 89%.

4. Discussion

In this study, Sudoscan was used in conjunction with AI algorithms to predict DAN, obtaining promising results. Some studies have used traditional clinical methods and biomarkers to assess the risk of neuropathy in diabetic patients [49]. These methods are valuable but often fall short of the accuracy and predictive power offered by advanced ML algorithms.

ML has been applied in various studies to predict some diabetic complications such as retinopathy, nephropathy, and cardiovascular diseases. However, research specifically focused on autonomic neuropathy remains very poorly explored. Hosseini Sarkhosh SM et al. [50] demonstrated that ML could enhance the predictive accuracy to 85% for diabetic nephropathy. Our study reinforces this finding, achieving high predictive accuracy with LR (92.6%) and RF (96.3%) classifiers and also with Lazy classifiers (97% with six models).

The advantages of our results are mainly the improved predictive accuracy compared to other studies, and the fact that Sudoscan, a non-invasive, fast, and friendly method, is used to assess the risk of neuropathy as opposed to other methods, such as nerve conduction or skin biopsies, could lead to the application of this method and their use in daily medical care. Integrating AI with Sudoscan data could lead to the development of automated real-time risk assessment tools. This is particularly beneficial in primary care settings where access to specialized diagnostic equipment may be limited. Our models could also be incorporated into electronic health record systems by providing alerts and recommendations to healthcare providers, enabling prompt interventions.

Traditional assessments for DAN, such as nerve conduction studies, are effective but invasive, costly both physically and materially, and not suitable for frequent monitoring. Our approach using Sudoscan and AI offers a less expensive alternative while maintaining high accuracy. Biochemical markers such as HbA1c, cholesterol, and triglycerides are useful for monitoring diabetes but these do not provide specific information about autonomic neuropathy. Combining these markers with Sudoscan data in our predictive models improves the overall assessment.

The study by Solatidehkordi Z. and Dhou S. [13], which involved 1275 subjects, is the only one that we could find that was comparable to ours and that attempts to identify early DAN in patients with diabetes using databases incorporating Sudoscan values. The machine learning models used in their study achieved a maximum accuracy of 82.82% with TabNet. By contrast, our study’s findings, across all examined models, are noticeably superior. Specifically, our study obtained an accuracy of 92.59% with LR, compared to 80.47% in the comparative study. Similarly, the RF model in our study obtained an accuracy of 96.30%, substantially higher than the 80.78% reported in the comparative study. These outcomes demonstrate how well the models we employed in our study performed. Most likely, the diversity of clinical data from our study is the primary element enhancing the performance of tested models. In contrast, the comparative study has less clinical detail despite having a larger dataset. Comprehensive clinical metrics, including blood test results (cholesterol, triglycerides, creatinine, and HbA1C), are included in our collection. The enhanced prediction performance of our models was probably influenced by the addition of these important medical criteria. These findings suggest that the prediction models tend to perform better not only with a larger number of samples but also with more detailed and medically relevant training data.

This study demonstrates promising results, but it is essential to validate these results in larger and more diverse populations to ensure method validation. Future research will focus on exploring and integrating additional variables and data sources to further improve predictive accuracy. Longitudinal studies could provide an effective long-term evaluation of our predictive models in the prevention of DAN and other diabetic complications. Monitoring values of interest over time may provide insights into the progression of neuropathy and the impact of early interventions.

5. Conclusions

This study represents a comprehensive analysis of Sudoscan’s ability to assess and predict diabetic autonomic neuropathy using statistical analysis and artificial intelligence using machine learning algorithms. The goal was to reduce the need for extra, more expensive, or even invasive testing by facilitating the initial diagnosis of neuropathy using baseline clinical data. RF had the best accuracy, at 96.3%, and had a mean accuracy achieved with k-fold cross-validation of 99.23%. The tested models’ performance demonstrates the promise of machine learning techniques for DAN prediction.

In conclusion, this study highlights the need to use Sudoscan as a diagnostic and prognostic tool in the management of diabetic neuropathy, but also that the integration of artificial intelligence algorithms would change and improve prevention and diagnostic strategies. These innovative methods can significantly contribute to changing the treatment of this disease, but also to a personalized approach to the disease.