To the Editor:
I read with interest the article by She et al, 1 “Exploration of a Multimodal Machine Learning Model Integrating Ultrasound and Clinical Indicators for the Diagnosis of Diabetic Peripheral Neuropathy.” The authors systematically developed and compared 4 machine learning models—logistic regression (LR), random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM)—for the diagnosis of diabetic peripheral neuropathy (DPN) using combined ultrasonographic and clinical data. Their finding that the RF model achieved superior predictive performance (AUC = 0.852) and identified key risk factors such as body mass index, 2‐hour postprandial C‐peptide, and diabetes duration represents a meaningful contribution to the field. I would like to offer 2 methodological considerations that may extend the discussion and inform future research in this area.
First, the study's reliance on a single‐center retrospective design with a modest sample size (235 patients) raises important considerations regarding model generalizability and stability. Although the authors employed 10‐fold cross‐validation and careful hyperparameter tuning to mitigate overfitting, the observed performance gap between the training and test sets suggests that the models may still be sensitive to the specific characteristics of the training data. In this context, methodological innovations from adjacent fields may offer valuable guidance. Long Short‐Term Memory networks with a later temporal attention mechanism can capture temporal dependencies in serological data, demonstrating that explicitly modeling time‐based patterns can improve predictive accuracy. 2 Translating such temporal modeling approaches to DPN research, where disease progression unfolds over years, could enhance the ability to capture dynamic risk trajectories rather than relying solely on cross‐sectional features. Furthermore, 3 proposed a 13‐layer residual neural network, showing that residual architectures can effectively capture complex, high‐dimensional relationships while maintaining computational efficiency. Given that DPN involves complex interactions among metabolic, inflammatory, and structural factors, residual network architectures may offer advantages over traditional ensemble methods in capturing these multifaceted relationships.
Second, the study's handling of ultrasonographic features warrants further consideration. The authors extracted multiple ultrasound parameters (width, thickness, cross‐sectional area, and shear wave velocity) from 2 nerve sites across 3 locations per nerve, resulting in a relatively high‐dimensional feature space relative to the sample size. While the RF model's built‐in feature importance mechanisms helped manage this dimensionality, the challenge of extracting meaningful information from high‐dimensional, potentially sparse data remains. In this regard, the singular pooling paradigm 4 offers an innovative solution. By replacing traditional global average pooling with a spectral pooling approach that preserves fine‐grained information from sparse feature maps, their method significantly improved classification performance in fetal ultrasound image analysis. Adapting such spectral decomposition techniques to DPN ultrasound data could enhance the extraction of subtle morphological and elastographic features that might otherwise be lost in conventional pooling or aggregation approaches. Additionally, the challenge of small‐sample DPN datasets may benefit from hybrid modeling approaches that balance dense representation with sparse variable selection. 5 Adapting such hybrid frameworks to DPN prediction could potentially improve model robustness and interpretability, particularly when integrating heterogeneous data types such as clinical variables and imaging features.
I commend the authors for their rigorous work and for advancing the application of machine learning in DPN diagnosis. Their findings underscore the promise of multimodal approaches and highlight important directions for future research. Addressing challenges related to sample size, temporal dynamics, and feature extraction will further strengthen the clinical utility of predictive models for DPN.
This work is supported by the Natural Science Foundation of Fujian Province (2023J05083).
Data Availability Statement
Data sharing not applicable to this article as no datasets were generated or analyzed during the current study.
References
- 1. She BY, Song M‐L, Chen W‐B, Wu K‐B, Lai Z‐H. Exploration of a multimodal machine learning model integrating ultrasound and clinical indicators for the diagnosis of diabetic peripheral neuropathy. J Ultrasound Med 2026; 45:1863–1876. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. Cai J, Li Y, Liu B, et al. Developing deep LSTMs with later temporal attention for predicting COVID‐19 severity, clinical outcome, and antibody level by screening serological indicators over time. IEEE J Biomed Health Inform 2024; 28:4204–4215. [DOI] [PubMed] [Google Scholar]
- 3. Cai J, Qi Y, Liu S, Chen R. A residual joint antenna network for joint transmit‐receive antenna subset selection in MIMO systems. IEEE Trans Antennas Propag 2025; 73:8366–8371. [Google Scholar]
- 4. Zhu S, Cai J, Xiong R, Zheng L, Ma D. Singular pooling: a spectral pooling paradigm for second‐trimester prenatal level II ultrasound standard fetal plane identification. IEEE Trans Circuits Syst Video Technol 2025; 35:12508–12523. [Google Scholar]
- 5. Cai J, Liu S, Wang B, Chen T, Zhu N, Chen R. Developing a CR‐SCAD algorithm for fibrosis and inflammatory activity analysis of chronic hepatitis C. Int J Mach Learn Cybern 2026; 17:13. [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
Data sharing not applicable to this article as no datasets were generated or analyzed during the current study.
