Personalized prediction of diabetic foot ulcer recurrence in elderly individuals using machine learning paradigms

2.1Previous studies on recurrence risk of diabetic foot ulcers

There have been several studies focusing on the recurrence risk of diabetic foot ulcers, but most of them have emphasized patients across all age groups. Studies often employed traditional statistical methods such as regression analysis to explore the influence of patients’ biological and clinical characteristics on recurrence risk [6].

2.2Application of machine learning in diabetes management

In recent years, significant progress has been made in the application of machine learning in the medical field. Some studies have attempted to use machine learning algorithms to predict the occurrence and recurrence of diabetic foot ulcers. These methods typically leverage clinical and biological data to identify patients’ risk factors [7]. However, research on elderly patients with diabetes is relatively scarce and requires more attention.

2.3Factors and diabetic foot ulcer recurrence in elderly patients

Our study specifically focuses on elderly patients with diabetes and considers a range of specific factors such as age, blood sugar control, alcohol consumption, smoking, etc., which may impact foot ulcer recurrence. Further research should delve into the complex relationships between these factors to gain a more comprehensive understanding of the recurrence risk in elderly patients with diabetes [8].

3.1Model selection

In this study, we opted for the Support Vector Machine (SVM) machine learning classifier to predict the recurrence risk of foot ulcers [9]. We then compared it with several common machine learning classifiers, including Support Vector Machine (SVM), XGBoost, KNN, Random Forest, and Decision Tree.

Support Vector Machine (SVM): SVM is a powerful classifier suitable for high-dimensional data and nonlinear relationships [10]. We selected SVM as the primary classifier due to its significant achievements in the medical field. We fine-tuned SVM hyperparameters, including C value, kernel type, and degree, to achieve optimal performance.

XGBoost: XGBoost is a gradient boosting algorithm suitable for large-scale data and efficient feature engineering [11]. We used XGBoost as a comparative model, optimizing its performance by adjusting parameters such as learning rate, tree depth, and the number of trees.

KNN (K-Nearest Neighbors): KNN is an instance-based algorithm, with predictions relying on the labels of previous instances most similar to the new instance [12]. We set K value and distance weights to explore its effectiveness in predicting foot ulcer recurrence risk.

Random Forest: Random Forest is an ensemble learning algorithm that combines multiple decision trees to improve predictive performance [13]. We set parameters such as the number of forests, tree depth, and others to ensure the model’s robustness and generalization ability.

Decision Tree: Decision trees are simple and intuitive classifiers helpful for understanding relationships between features. We set parameters like maximum depth and minimum leaf node number to balance model complexity and performance.

3.2Data preprocessing

During data preprocessing, we initiated by cleansing the data, removing outliers, duplicates, and missing values. For missing values, we employed appropriate imputation methods such as mean or median filling to ensure data integrity.

Next, we conducted feature engineering, binary encoding discrete data for model understanding. For continuous data, we utilized the IQR method for outlier screening to enhance the model’s robustness.

Figure 2.

Data preprocessing workflow.

Lastly, to better comprehend the relationship between data features and foot ulcer recurrence, we performed correlation analysis, plotting correlation matrices to assist in selecting highly correlated features with the target variable for integration. The overall workflow is depicted in Fig. 2.

3.3Genetic factors

In order to enhance the comprehensiveness of our predictive model, we recognize the importance of incorporating genetic factors into the analysis. Genetic markers can potentially play a crucial role in predicting the recurrence of diabetic foot ulcers (DFUs) in elderly diabetic patients, contributing to a more holistic understanding of the risk factors involved.

As part of our data collection process, we propose the inclusion of genetic data from study participants. This may involve obtaining information on specific genetic markers associated with diabetes and its complications. Collaborating with geneticists or utilizing existing genetic databases could provide valuable insights into the genetic predisposition to DFU recurrence.

Once genetic data is collected, it should be integrated into our machine learning models alongside other high-risk factors, such as age, blood sugar control, alcohol consumption, and smoking. The combined analysis of clinical and genetic data aims to improve the accuracy and comprehensiveness of our predictive model.

This addition not only strengthens the predictive power of our model but also contributes to advancing our understanding of the complex interplay between genetics and diabetic foot ulcer recurrence in elderly individuals.

3.4Model evaluation

To assess the performance of the models, we employed common classification metrics, including Recall, Precision, and Accuracy [14]. These metrics provide a comprehensive understanding of the model’s performance in various aspects.

During the experimental process, we utilized the K-fold cross-validation algorithm, dividing the dataset into multiple subsets to train and validate the model more comprehensively. This helps prevent overfitting and enhances the model’s generalization ability.

Through the mentioned methods and experimental design, our goal is to establish an accurate, robust, and interpretable predictive model for the recurrence risk of foot ulcers.

4.1Data collection and processing

We collected high-risk factors for ulcer recurrence from 138 elderly diabetic patients, including gender, age over 70, thrombosis, alcohol consumption, smoking, strenuous physical labor, surgical intervention, blood sugar control, use of hormones and immunosuppressants, wound occurrence time, wound size (in CM), wound location, compression stockings, and BMI.

Raw data often contains missing values (NA or blanks), making it unsuitable for direct training. Additionally, the presence of outliers (values beyond the normal range) can negatively impact model performance [15]. Effective data processing is fundamental to improving predictive accuracy.

In data processing, we initially conducted data cleansing, removing outliers, duplicates, and handling missing values through deletion or imputation. Subsequently, we performed a correlation analysis on the cleaned data, as shown in Fig. 3. Highly correlated data were fused to reduce fitting complexity.

Figure 3.

Correlation analysis of high-risk factors for ulcer recurrence data.

Afterwards, the data was categorized into continuous and discrete variables based on their value ranges. For discrete variables, we transformed them into binary encoding. For continuous variables, we initially applied the IQR method for outlier detection. The distribution of continuous data is illustrated in Fig. 4. Upon the identification of potential outliers, a cautious and iterative process was undertaken. Outliers were carefully examined to determine the validity of their values. We opted for a pragmatic approach, considering domain knowledge and potential impact on the overall dataset. Outliers were then either corrected through appropriate imputation methods or, in cases where their values were deemed erroneous, they were removed from the dataset.

Following the removal of outliers, we subjected the continuous data to scaling procedures. Feature scaling aims to mitigate the impact of feature scales on the model, thereby enhancing accuracy [16]. Additionally, for a more comprehensive examination of the relationship between features and ulcer occurrences, we integrated feature data with ulcer incidents for visualization, as depicted in Figs 5 and 6.

Figure 4.

Distribution of continuous high-risk factors for ulcer recurrence data.

Figure 5.

Integration of feature data and ulcer: Continuous feature fusion and target distribution.

Figure 6.

Integration of feature data and ulcer: Stacked bar chart of categorical features and target.

4.2Experimental procedure

In this study, Support Vector Machine (SVM) is employed as the predictive classifier, and it is compared with XGBoost, KNN, Random Forest, and Decision Tree. To ensure a fair and accurate comparison, each algorithm undergoes rigorous parameter tuning, aiming to optimize its performance. Table 1 provides a detailed overview of the model parameters for each classifier, showcasing the specific configurations chosen during the tuning process. The careful adjustment of parameters for each algorithm aims to capture the nuances of their individual strengths and characteristics, contributing to a robust and unbiased comparative analysis. This approach ensures that each algorithm is equipped with optimal settings, fostering a comprehensive evaluation of their effectiveness in predicting the recurrence risk of diabetic foot ulcers (DFUs) in elderly diabetic patients.

Accuracy: Summarizing the Performance of the Classification Task and assessing the count of accurately predicted instances among all data instances.

Recall: The proportion of participants with ulcer attacks and correctly classified among all participants.

Precision: How many people identified as having an ulcer attack belong to this category.

The formulas for calculating Recall, Precision, and Accuracy are as follows:

Here, TP stands for true positive, TN for true negative, FP for false positive, and FN for false negative. After preprocessing and standardization, we split the dataset into training and testing sets with a ratio of 7:3. Subsequently, the K-fold cross-validation algorithm is applied to train the model. The training dataset is evenly divided into k subsets, treating each subset as a validation fold once. Cross-validation is repeated k times, each time training the model with the remaining folds and validating the model’s performance on the respective validation fold for training optimization. In our experiment, we set k to the empirical value of 5, striking a balance between computation and accuracy. Finally, the accuracy of the model on the test set is used as the ultimate performance metric.

4.3Experimental results

We divided the data into training and validation sets at a ratio of 7:3. The validation set was excluded from data processing and model training, serving only for the final validation of model performance. The performance of the model on the validation set reflects its generalization ability. Model evaluation is based on the model’s performance on the validation set. The final results are presented in Table 2.

From Table 2, it can be observed that SVM achieved the highest accuracy at 93%, while Decision Tree exhibited the lowest performance with only 79%.

4.4Interpretability of machine learning models

Our analysis reveals that specific features play a pivotal role in the predictive outcomes of our model. Notably, surgical intervention demonstrate significant influence, shedding light on the critical decision-making factors behind our predictions.

We outline the decision process of our machine learning models, emphasizing the sequential steps taken in making predictions. Understanding the pathway to each prediction enhances transparency and aids healthcare professionals in grasping the rationale behind the model’s decisions.

This section provides a user-friendly interpretation of the model’s predictions, making it accessible and useful for healthcare professionals and patients alike.In addition to our current approach, we consider the exploration of alternative interpretable machine learning methods or post-hoc interpretability techniques.