Globally, diabetes mellitus (DM) has significantly impacted healthcare costs and socioeconomic burdens, escalating from 966 billion United States dollars (USD) in 2021 to a projected 1,054 billion USD by 2045. A recent epidemic report estimated the global prevalence of diabetes at approximately 10.5%, set to rise to 12.2% by 2045 (Sun et al., 2022). Similarly, a recent epidemiological report indicated that the prevalence of chronic kidney disease (CKD) was 10.0% in adult populations globally; however, this value may be underestimated. Akin to DM-related impacts, CKD-related healthcare stands as the primary driver of medical and social costs in most countries (Sundström et al., 2022). Patients with severe DM often progress to CKD, which can impose greater care challenges; hence, CKD could be considered a more severe condition in patients with DM (Parving et al., 2006; Thomas et al., 2016). To monitor DM progression and prevent the development of severe disease, a low-cost, quick, and noninvasive method is needed.

Individuals with DM and CKD commonly experience peripheral nerve disorders (Baumgaertel et al., 2014), particularly peripheral neuropathy (Pop-Busui et al., 2017; Ezzeldin et al., 2019; Feldman et al., 2019; Karlsson et al., 2019). Clinical symptoms associated with neuropathies include pain, impaired thermal discrimination, sensory deficits, reduced motor function, and diminished or absent distal reflexes (Callaghan et al., 2015; Pop-Busui et al., 2017; Ezzeldin et al., 2019; Feldman et al., 2019). These symptoms significantly impact daily activities and could be critical in pre-DM and early stages of CKD (Singleton et al., 2001; Ziegler et al., 2008; Im et al., 2012; Smith and Singleton, 2012; Bongaerts et al., 2013; Asghar et al., 2014; Moorthi et al., 2019). Monitoring changes in sensorimotor function, an evident neurological feature, might be a feasible strategy to monitor the progression of both diseases. Early detection of these diseases could aid patients and clinicians in comprehending disease progression and subsequently achieving improved prognosis (Tesfaye et al., 2010; Bernardi et al., 2011; Bril et al., 2011; Dyck et al., 2011; Kempler et al., 2011; Spallone et al., 2011; Finnerup et al., 2015; Pop-Busui et al., 2017). Despite easily identifying neuropathies in both groups, the mechanism of neuropathies in DM and CKD remains unclear and may result in different neuropathic symptoms or sensorimotor features in these two diseases (Biessels et al., 2014; O Brien et al., 2014; Vincent et al., 2011; Zenker et al., 2013). For example, patients with CKD were found to exhibit poorer light touch sensory function than non-CKD participants, even after excluding the effects of DM (Moorthi et al., 2019). In other words, sensorimotor performance might vary between DM and CKD, despite both having neuropathies as diagnoses. Due to potential differences in sensorimotor performance between DM and CKD, evaluating sensorimotor function to monitor disease progression could prove to be a valuable, low-cost method.

Recent studies introduced a novel pinch-holding-up-activity (PHUA) test, using sensorimotor function measurements with robust psychometric properties (Chiu et al., 2009). These investigations aimed to discern disparities in hand sensorimotor performance hand between patients with DM and healthy adults (Chiu et al., 2014; Hsu et al., 2015), as well as differences between patients with CKD and healthy adults (Tu et al., 2019). These studies showed significant differences in sensorimotor performance between healthy adults and patients with peripheral neuropathic hands. Sensorimotor parameters—such as force ratio and percentage of maximal pinch force—were notably larger in patients with neuropathy, indicating that the use of inefficient or improper hand performance strategies. Furthermore, these parameters displayed medium-to-high correlations between sensory conditions and fine motor function (Shieh et al., 2011; Hsu et al., 2013).

However, these studies primarily relied on a limited set of parameters—such as force ratio or maximal pinch ratio—derived from specific events within the signals to determine the inferior sensorimotor performance in individuals with neuropathic hands. Unfortunately, these parameters proved insufficient for distinguishing sensorimotor features between the DM and CKD groups using current analytical approaches.

In recent years, machine learning has rapidly developed for human motion analysis. For fundamental research, Liu et al. proposed several base studies for extracting interpretable and explainable features to help build machine learning models for human activity recognition (HAR) (Liu et al., 2021; Hartmann et al., 2022; Liu et al., 2023) and published a feature extraction library for time-series data (Barandas et al., 2020). Hartmann et al. also found that high-level and interpretable features can be used in few-shot learning, and the results were promising (Hartmann et al., 2023). At the same time, a new branch in machine learning, called deep learning, has found widespread use in complex and noisy signal applications, where conventional analyses might struggle to extract pertinent information (LeCun et al., 2015; Goodfellow et al., 2016). Ideal deep learning models use original data or parameters without preprocessing or human-selected procedures. Previous studies have demonstrated the efficacy of convolutional neural networks (CNN) in appropriately handling time-series images or signals of human motion (Hannink et al., 2016; Hannink et al., 2017; Kautz et al., 2017). Architectures like VGG (Hannink et al., 2016; Hannink et al., 2017; Kautz et al., 2017), ResNet (Wang et al., 2017; Cheng et al., 2021), and dilation CNN (Arık et al., 2017; Bai et al., 2018; Lei et al., 2019) have emerged as strong candidates for processing time-series signals. Despite their demonstrated capabilities in processing time-series data, current studies rarely classify different peripheral neuropathies based on sensorimotor features of hand performance. Therefore, this study aimed to develop three distinct CNN models for classifying DM and CKD diagnoses based on hand sensorimotor function.

The model parameters used to demonstrate the model efficiency are listed in Table 2. In the FCN model, the total parameters employed were 502,979, whereas dil-CNN, shallower in structure, used 205,123 parameters. In ResNet, the total parameters is 379,907 (Table 2). The model structures of the FCN and dil-CNN were similar; however, the FCN was deeper, with two additional convolutional layers. The FCN and dil-CNN differed in its dilation rate, increasing from 2 to 4 in the seven–10 convolutional layers (Figures 3A, B). ResNet, following the structure of the original study with two blocks (He et al., 2016) repeated four times (Figure 3C). To mitigate overfitting risks, our models were designed with reference to the original article were not replicated at the same depth, such as VGG-19 or ResNet-34, we opted for a shallower architecture during the validation phase, refining it iteratively through trial and error.

The final test results indicate intra-participant accuracies for FCN, ResNet, and dil-CNN as 0.926, 0.874, and 0.953, respectively. Meanwhile, inter-participant accuracies for FCN, ResNet, and dil-CNN stand at 0.879, 0.875, and 0.898, respectively (Table 3). The weighted F1 scores (wF1) for intra-participant assessments of these models were 0.927, 0.873, and 0.953, whereas the inter-participant wF1 scores were 0.877, 0.869, and 0.897, respectively (Table 3).

The evaluation of the model performance using the AUC is shown in Table 4; Figure 4. Across the intra-participant dataset, the AUCs for healthy adults were 0.984, 0.954, and 0.993 for FCN, ResNet, and dil-CNN, respectively. In the inter-participant dataset, these AUCs were 0.968, 0.966, and 0.966, respectively. The AUCs of the three models for DM in the intra-participant dataset were 0.983, 0.961, and 0.994, respectively. Assessing DM within the interparticipant dataset yielded AUCs of 0.972, 0.952, and 0.973. Finally, for CKD in the intra-participant dataset, AUCs were 0.995, 0.938, and 0.999. For the inter-participant dataset, the AUCs of the three models were 0.953, 0.916, and 0.977, respectively (Table 4).

The contemporary diagnosis of DM and CKD in clinical settings relies on blood tests (Harris and Eastman, 2000; Stevens and Levey, 2009), considered the golden standard. However, the proposed PHUA test and the findings in this study do not aim to replace this gold standard in clinical practice. Instead, the PHUA test, combined with a deep learning model, offers a quicker and less invasive method for distinguishing between patients with DM and CKD in medical scenarios that require rapid screening or in places where laboratory examinations are not readily available. Our findings suggest that PHUA with a deep learning model could potentially discern differences in the sensorimotor features of the hand between CKD and DM. Therefore, this study may suggest a simple test to aid in consistently monitoring the progression of DM and preventing its advancement to CKD (Parving et al., 2006; Thomas et al., 2016). Previous studies demonstrated that regular follow-up examinations can reduce the severity of complications, particularly for individuals >45 or <45 but with significant risk factors such as obesity and a family history of DM (Pippitt et al., 2016). Assessing sensorimotor capability is typically simple and rapid in clinical, community, or home-based scenarios. To mitigate evaluation difficulties and potential inter-tester errors, the PHUA could serve a suitable apparatus for clinical assessments due to its well-define design rationale. Its design was developed to challenge participant reflex motor responses while performing upward movements by pinching a glossy surface (Chiu et al., 2009; Hsu et al., 2009; Shieh et al., 2011). The findings of this study indicate that the PHUA test, coupled with a deep learning model, could serve as a potential tool to evaluate the sensorimotor function of the hands, and differentiate between DM and CKD based on sensorimotor impairments, owing its high accuracy and AUC. Given its higher accuracy, our proposed model with the PHUA test could be a viable option for swift clinical screening and monitoring, particularly for subjects who may be neglecting risks while being away from medical providers.

Neuropathy in both DM and CKD typically manifests as neural system damage that impairs sensorimotor performance (Tesfaye et al., 2010; Baumgaertel et al., 2014), but research on distinctions between DM and CKD in this context is limited. Our proposed models suggest that deep learning models can uncover variations in sensorimotor patterns between DM and CKD (Figure 5) and marked the sample numbers in Figure 5 because of the imbalanced sample size between different groups (Jannat et al., 2023). The confusion matrix generated by the dil-CNN model implies potential differences in PHUA-based sensorimotor performance between DM and CKD, aligning with previous findings (Moorthi et al., 2019). Earlier research demonstrated that patients with CKD exhibit poorer light touch sensation compared with healthy adults and those with DM (Moorthi et al., 2019). However, the impact of worsened sensory sensation on clinical evaluation of sensorimotor patterns remains unexplored. The study findings indicate that a robust model could discern specific differences in sensorimotor features between DM and CKD. Despite similarities in impaired sensorimotor function—such as sensory impairment and motor function deficit—between DM and CKD, this discovery suggests that the decline in sensorimotor function in these conditions might arise from different mechanisms or conditions. However, although the deep learning model highlights differences in sensorimotor patterns between DM and CKD, it remains a black box, unable to specify the exact degenerative processes causing these differences (Castelvecchi, 2016). Notably, sensory or motor function impairment affects PHUA performance (Chiu et al., 2009; Shieh et al., 2011; Hsu et al., 2012; Hsu et al., 2013). Previous studies have suggested that neuropathies and the underlying mechanisms causing sensorimotor degeneration in individuals with DM or CKD may differ (Tesfaye et al., 2010; Callaghan et al., 2015; Pop-Busui et al., 2017; Ezzeldin et al., 2019; Feldman et al., 2019; Karlsson et al., 2019). Neuropathy in DM stems from uncontrolled blood glucose levels, deforming blood vessels and leading to insufficient neuronal nourishment (Tabatabaei-Malazy et al., 2011; Baumgaertel et al., 2014). Damage primarily occurs in the distal body parts (Ezzeldin et al., 2019). In contrast, neuropathy in CKD results from toxic substances in the blood (Baumgaertel et al., 2014; Arnold et al., 2016), affecting nervous tissues throughout the body, including the muscles, neural system, and metabolic system (Arnold et al., 2016). Although these models solely differentiate motor patterns between DM and CKD, prior research suggests that differing neuropathic mechanisms also imply varying severity levels and affected the body system ranges (Moorthi et al., 2019).

During the model design and adjustment phase, a smaller kernel size resulted in higher accuracy within the modified model structure. This finding aligns with previous research suggesting that smaller kernels may lead to better convergence, a principle noted in the original VGG designer work on image recognition (Simonyan and Zisserman, 2014). The three proposed model architectures suffer from overfitting. During the design phase, regularization to prevent overfitting, such as the batch norm layer, L1 or L2 regularization (Krogh and Hertz, 1992; Schmidhuber, 2015), and dropout (Srivastava et al., 2014), was tested, and the overfitting condition did not improve. Moreover, ResNet displayed severe overfitting compared with the other dil-CNN models, but with more parameters included in the model (Table 2). This could be attributed to the shortcut structure of ResNet, as demonstrated in previous research where a neural network with shortcuts can fit signals effectively given enough layers (Lin and Jegelka, 2018). Although ResNet structure in this study was not extremely deep, overfitting persisted. Attempts were made to explore models with fewer layers, but they resulted in underfitting compared with the final proposed structure. Further research on shortcut-designed models for time-series sensorimotor performance data is warranted to address these challenges.

To accommodate the increased computational power and memory requirements of larger models, the final model included only six convolutional layers in the FCN and dil-CNN, whereas six identity blocks were used in ResNet. Notably, the last layer of average pooling outperformed the flattened layer (Lin et al., 2013). No additional dense layer followed global average pooling, a strategy that mitigated overfitting and increased testing accuracy. This aligns with the findings in deep learning model design for image recognition post-2013 (Lin et al., 2013; He et al., 2016; Szegedy et al., 2016).

The dil-CNN showed superior robustness with fewest parameters among compared models. The total number of parameters in the dil-CNN was 205,123, nearly half that of ResNet (Table 2) and one-third that of the FCN. The dil-CNN had a higher accuracy than the FCN (0.953 vs. 0.926 in intra-participant; 0.898 vs. 0.879 in inter-participant; Table 3). In the proposed model, dil-CNN exhibited the best performance. Leveraging the design of dilation in convolution, the dil-CNN effectively combines information across extended time-series data. In our study, the structure based on VGG remains unchanged, with dilation being the sole modification, yet it notably enhances accuracy. However, we did not evaluate the dilation kernel with a residual network (Bai et al., 2018) in our proposed models because of observed overfitting in ResNet. Future research should explore additional adjustments to the dil-CNN, potentially incorporating a shortcut connection. Nevertheless, for classification tasks involving sensorimotor signals, the dilation CNN might prove superior to other structures.

Previous studies have relied on several human-selected parameters, resulting in a lower AUC for patients with DM and healthy adults (AUC = 0.724) (Chiu et al., 2014), as well as for CKD and healthy adults (AUC = 0.848) (Tu et al., 2019), whereas the dil-CNN exhibited higher AUC values (AUC > 0.96, Table 4). Notably, complete raw data were included in our analysis, contributing to improved prediction accuracy and potential for handling multiple classes using deep-learning models, as indicated by our results.

Our study has several limitations that warrant acknowledgement. All proposed models in this study faced challenges related to overfitting, despite incorporating regularization and batch-normalization layers. This issue might stem from the restricted total number of participants, suggesting the need for a larger sample size, particularly for machine-learning models reliant on deep learning. Increasing the sample size could enhance model stability and mitigate overfitting concerns in real-world applications. In addition, few-shot learning with pre-processing and feature engineering (Hartmann et al., 2023) could be a possible solution for PHUA testing, especially using the model in following subjects with a high potential risk of DM and CKD. Moreover, the models based on deep learning represent black boxes in this study, lacking explicit explanations for distinguishing sensorimotor performance in DM from that in CKD. Although visualization methods such as Grad-CAM (Selvaraju et al., 2017), Grad-CAM++ (Chattopadhay et al., 2018), and I-GOS (Qi et al., 2019) exist to potentially elucidate model attention, these approaches offer visual insights without statistical significance and are susceptible to misleading interpretations (Subramanya et al., 2019). Developing interpretable AI warrants further investigation, involving visualization techniques like Grad-CAM, sensitivity analysis, and generative models. In the other way, human-selected features with interpretable feature engineering can be an innovative method for developing a explainable machine learning model (Hartmann et al., 2022; Hartmann et al., 2023). These efforts aim to assist clinical staff identifying meaningful features or key points for evaluating patient sensorimotor performance. Another limitation lies in the insufficient variety of diagnoses considered, limiting the model adaptability in clinical scenarios. Future studies should encompass diagnoses with similar sensorimotor impairments, such as carpal tunnel syndrome, peripheral neural damage, or neuromuscular disorders, to refine the model output accuracy. An additional limitation is that our study represents just the initial step toward rapid screening using PHUA with a DNN. Furthermore, participants with DM and CKD were exclusively recruited at severe stages. To establish a comprehensive clinical solution for rapid screening or sensorimotor evaluation, different stages and progressions of DM and CKD should be considered, ensuring model robustness across diverse patient profiles.

Our findings underscored the potential of PHUA coupled with a deep learning model to differentiate various sensorimotor patterns among healthy adults and patients with DM and CKD. For the evaluation of human sensorimotor performance based on time-series signals, the dilated CNN structure demonstrated notable accuracy and efficiency. Future studies require larger sample sizes encompassing varying disease severities and considerations for the comorbidity of DM and CKD to advance the next-generation model. Developing an interpretable model is crucial to facilitate its practical application in clinical settings.

In conclusion, our study highlights the capacity of the DNN model to distinguish between healthy adults, participants with DM, and CKD participants through the innovative motor performance evaluation tool, PHUA. PHUA integrated with the dil-CNN model exhibits remarkable stability and accuracy, presenting sensorimotor performance assessment as a novel approach to aid in evaluating CKD and DM diagnosis stages and offering an effective screening method for both disorders. This study presents an innovative application of machine learning in clinical evaluation, particularly for patients with DM and CKD.