Risk Factors in Males and Females for Disease Classification Based on International Classification of Diseases, 10th Revision Codes ^†

Abstract

We developed a machine learning model for disease classification based on the International Classification of Diseases, 10th Revision (ICD-10) codes, analyzing male and female groups using seven features. The three most prevalent ICD-10 classes covered over 98% of the data. Features were selected using the least absolute shrinkage and selection operator, ridge, and elastic net, followed by the mean decrease in accuracy and impurity. A random forest classifier with five-fold cross-validation showed improved performance with more features. Using Shapley additive explanations, age, BMI, respiratory rate, and body temperature were identified as key predictors, with gender-specific variations. Integrating gender-specific insights into predictive modeling supports personalized medicine and enhances early diagnosis and healthcare resource allocation.

1. Introduction

Gender and gender differences play a critical role in the development and application of machine learning models for disease classification. These differences significantly impact the epidemiology, clinical manifestation, and progression of various medical conditions, including dermatological disorders, autoimmune diseases, gender-specific cancers, and other illnesses documented in hospital records based on the International Classification of Diseases, 10th Revision (ICD-10) classification system.

Machine learning (ML) has emerged as a transformative tool in the medical field, significantly advancing disease classification and predictive analytics. Various ML algorithms, including Decision Trees (DTs), Support Vector Machine (SVMs), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Artificial Neural Networks (ANNs), have been widely applied to analyze complex medical datasets and predict disease outcomes with high accuracy [1,2,3]. The effectiveness of classification models is highly dependent on selecting relevant features, particularly when handling high-dimensional medical data, making feature selection crucial for improving interpretability, reducing computational complexity, and mitigating overfitting. Regularization techniques, such as the least absolute shrinkage and selection operator (LASSO), ridge regression, and elastic net, are widely employed to identify informative features while addressing multicollinearity issues [4,5,6]. These methods impose penalties on model coefficients, ensuring sparsity and enhancing generalizability. Beyond regularization-based feature selection, feature importance ranking techniques, such as the mean decrease in accuracy (MDA) and the mean decrease in impurity (MDI), are used to refine classification models by evaluating the predictive contribution of each variable [7,8,9,10]. MDA is used to assess feature significance by measuring classification accuracy reduction upon feature exclusion, while MDI is used to quantify a feature’s contribution to decision tree-based models through impurity reduction across splits. Integrating these techniques with ICD-10 classification facilitates precise disease identification, enabling early detection, optimized resource allocation, and evidence-based policy-making and making ML classification easier for both analysis and further implementation. In the ML model’s classification and performance, the feature selection is important. In this study, hybrid feature selections based on L1 (LASSO) and L2 (Ridge) regularization with MDA and MDI were conducted to investigate the implementation of machine learning models based on a constrained set of features.

2. Disease Classification

The model in this study was developed based on the data collected from hospital records. Data were pre-processed to deal with missing values and standardize data for robust analysis. The RF model and its hyper-parameter tuning were adopted to predict disease categories. Model performance was evaluated, and the feature importance was analyzed by using Shapley additive explanation (SHAP).

2.1. Data Collection

Ethical approval was acquired under an exemption from the Rangsit University Ethics Committee, under the Helsinki Declaration. Secondary ICD-10 data were collected from Nong Han Hospital, Udon Thani, Thailand, from 2020 to 2022.

2.1.1. ICD-10 Classification

ICD-10 provides a standardized system for categorizing diseases, health conditions, and procedures [11,12,13,14]. Being widely used in clinical documentation and research, it features alphanumeric codes that support consistent data analysis across healthcare systems. We used codes A00–A99, covering infectious and parasitic diseases of significant public health importance.

2.1.2. Data Summary

Focusing on the response variable, the distribution of intestinal infectious diseases (A00–A09), tuberculosis (A15–A19), and other bacterial diseases (A30–A49) between males and females were compared. Females showed a higher proportion of intestinal infectious diseases (72.81%) than males (50.13%). Conversely, tuberculosis (26.13%) and other bacterial diseases (23.74%) were more prevalent in males (14.81%) than in females (12.38%), respectively, showing gender-specific differences in disease susceptibility.

2.2. Data Pre-Processing

Before analysis, data preprocessing was conducted for data cleansing, transformation, and imputation to ensure consistency and completeness. Three major disease classes were used: intestinal infectious diseases, tuberculosis, and other bacterial diseases. Intestinal infectious diseases were more prevalent among females (72.81%) than males (50.13%), whereas tuberculosis and other bacterial diseases were more frequent in males (26.13 and 23.74%) than in females (14.81 and 12.38%). For model training, the dataset was split into the training set (80%) and the testing dataset (20%). A five-fold cross-validation strategy was applied, where the training set was partitioned into five subsets to enhance model generalizability. To minimize class imbalance, class weights were assigned inversely proportional to class frequencies using

$$w_i={\displaystyle \frac{N}{k\times n_i}}$$

(1)

where $N$ is the total sample size, $k$ is the number of classes, and $n_i$ represents class size [4,5,6]. This weighting approach mitigates bias toward majority classes, improving predictive performance for minority classes. The class weights for the male and the female across three disease groups were evaluated as shown in

Table 1. Frequency and class weights of three major disease classes across different genders.

Class	Code Range	Description	Males			Females
			Frequency	Percentage	Weight	Frequency	Percentage	Weight
1	A00–A09	Intestinal infectious diseases	2764	50.13	0.67	4793	72.81	0.46
2	A15–A19	Tuberculosis	1441	26.13	1.28	975	14.81	2.25
3	A30–A49	Other bacterial diseases	1309	23.74	1.40	815	12.38	2.69
		Total	5514	100.00		6583	100.00

.

2.3. Feature Selection Techniques

Penalized regression techniques, such as LASSO, ridge, and elastic net, are widely used for feature selection to improve classification performance in disease prediction and biomedical applications [15,16,17]. As the response variable has three classes, the multinomial logistic regression (MLR) was adjusted. MLR was used to model the probability of each class $k$ using the following softmax function.

$$P(y_i=k|x_i)={\displaystyle \frac{exp\left({\beta }_0+{\beta }_k{x}_i\right)}{{{\displaystyle \sum }}_{j=1}^{K}exp\left({\beta }_0+{\beta }_j{x}_i\right)}}$$

(2)

where $y_i$ is the class label for sample $i$ taking values in {1, 2, 3}, $x_i$ is the feature vector for sample $i$, and ${\beta }_k$ is the coefficient vector corresponding to class $k.$ The standard objective function (negative log-likelihood) for multinomial logistic regression is defined by

$$L\left(\beta \right)=-{\displaystyle \sum }_{i=1}^n{\displaystyle \sum }_{k=1}^{K}I\left(y_i=k\right)logP(y_i=k|x_i)$$

(3)

where $I\left(y_i=k\right)$ denotes an indicator function that equals to 1 if $y_i=k$ and 0 otherwise. Subsequently, LASSO applies L₁ regularization, enforcing sparsity so that Equation (3) is adjusted to

$$L_1\left(\beta \right)=L\left(\beta \right)+{\lambda }_1{\displaystyle \sum }_{k=1}^K{\displaystyle \sum }_{j=1}^p\left|{\beta }_{j,k}\right|$$

(4)

where ${\lambda }_1$ is the L₁ regularization parameter. The second term enforces sparsity, making some coefficients exactly zero. Ridge applies L₂ regularization to prevent large coefficients. Equation (3) is then modified to

$$L_2\left(\beta \right)=L\left(\beta \right)+{\lambda }_2{\displaystyle \sum }_{k=1}^K{\displaystyle \sum }_{j=1}^p{\beta }_{j,k}^2$$

(5)

The L₂ penalty prevents large values in β but does not enforce sparsity. Combining L₁ and L₂ regularization, the elastic net is newly defined as

$$L_{1,2}\left(\beta \right)=L\left(\beta \right)+{\lambda }_1{\displaystyle \sum }_{k=1}^K{\displaystyle \sum }_{j=1}^p\left|{\beta }_{j,k}\right|+{\lambda }_2{\displaystyle \sum }_{k=1}^K{\displaystyle \sum }_{j=1}^p{\beta }_{j,k}^2$$

(6)

where ${\lambda }_1$ and ${\lambda }_2$ control sparsity and shrinkage, respectively. To observe the importance of each feature of males and females, the feature importance score including MDA and MDI was applied to the training set of 80% [8,9,10]. The conceptual idea is described as follows.

• MDA is a feature importance measure used in machine learning, particularly in RF. It quantifies the impact of each predictor by computing the decrease in model accuracy when the values of that feature are randomly permuted. A higher MDA score indicates a more influential feature so it is widely applied in classification and regression tasks.
• MDI is a feature importance metric used in tree-based models, particularly RF. It measures the reduction in impurity (e.g., Gini index or entropy) contributed by each feature across all decision trees. A higher MDI score indicates a more influential feature. However, MDI is biased towards variables with more categories or higher cardinality, necessitating careful interpretation in feature selection and model evaluation.

2.4. Classification Models and Performance

Regarding the ICD-10 disease classification based on the dataset, we defined three major classes (Table 1). An RF classifier, one of the most effective ML models, was implemented. Per-parameter tuning and evaluation metrics of RF are described as follows.

RF was the main classification model used in this study. It is an ensemble-based machine learning approach frequently employed for classifying infectious and parasitic diseases using ICD-10 codes [2,18,19]. RF generates multiple decision trees during the training phase and aggregates their predictions to enhance accuracy and mitigate overfitting. By constructing each tree from randomly selected subsets of features and samples, RF improves model robustness against noise and imbalanced data.

Grid search strategy was used to systematically evaluate all possible combinations of parameters within a given parameter set, providing a comprehensive exploration of the parameter space [20]. The optimal parameters identified for the model for the male included the maximum depth of 15, minimum samples split of 15, and 60 estimators, while the best parameters for the female model included the maximum depth of 15, minimum samples split of 5, and 100 estimators.

Feature importance provides information on the key predictors contributing to the model. Class weighting is used to mitigate data imbalance by ensuring equitable representation of all classes. To evaluate the impact of these factors on model performance, the 3 × 3 confusion matrix for ICD-10 disease classification was used to derive essential metrics, including accuracy, precision, recall (sensitivity), specificity, F1-score, balanced accuracy, and the area under the curve (AUC) [10,21,22,23]. The F1-score is used to balance precision and recall, whereas AUC is used to assess model performance across various thresholds. Given the presence of three disease classes, all metrics were averaged and subsequently reported to provide a comprehensive evaluation in this study.

2.5. SHAP Analysis

SHAP is widely used in interpreting machine learning models by assessing feature importance through Shapley values, derived from cooperative game theory [1,24,25]. It is used to quantify each feature’s contribution to the model’s predictions, offering information on model behavior. For classifying infectious and parasitic diseases using ICD-10 codes, key diagnostic factors were used in SHAP to improve interpretability and healthcare decision-making in this study. Generally, SHAP analysis is applied to the best-performing model to ensure accurate feature attribution.

3. Results

Python (version 3.10.9, conda-forge distribution) was used for all analyses. Core libraries included pandas (v1.5.2) for data manipulation, SciPy (v1.9.3) and statsmodels (v0.13.2) for statistical procedures, scikit-learn (v1.6.1) for feature selection and machine learning models, and PyTorch (v2.1.2+cu118) for computational tasks. These tools supported model implementation, feature evaluation, and performance analysis. Gender-specific feature sets were evaluated to identify the most suitable features for optimal classification. The models were assessed using cumulative feature inclusion strategies. Finally, SHAP analysis was conducted to interpret the contribution of individual features and understand the influence of each feature on classification outcomes and the model’s decision-making process.

3.1. Feature Selection Results

The feature selection results from LASSO, Ridge, and Elastic Net regression highlighted variations in gender-specific disease classification. Age, temperature, respiratory rate (RR), pulse, and BMI were selected for both males and females, while occupation and status exhibited lower coefficients. Then, the MDA and MDI scores were evaluated for all significant features in both males and females (

Table 2. Feature coefficients from different penalized regressions.

No.	Feature	LASSO		Ridge		Elastic Net
		Males	Female	Male	Female	Male	Female
1	Age	0.3838	0.2013	0.3936	0.2107	0.3879	0.2059
2	Temperature	0.1223	0.1578	0.1257	0.1606	0.1252	0.1596
3	Pulse	0.0884	0.0995	0.1025	0.1139	0.0991	0.1112
4	RR	0.0799	0.0901	0.0880	0.0975	0.0867	0.0964
7	BMI	0.0688	0.0113	0.0804	0.0221	0.0787	0.0206
6	Marital Status	-	0.0345	-	0.0373	-	0.0392
7	Occupation	−0.0130	−0.0200	−0.0213	−0.0277	−0.0222	−0.0270

).

In Figure 1, feature importance scores of MDA and MDI highlighted gender-specific differences. For MDA, age and temperature have the highest scores for males, while age and BMI were important for females. For MDI, age and BMI dominated for both genders, with pulse important for females. Setting the threshold value of 0.01, MDA showed five features for both males and females, while MDI selected six features for males and seven for females (Status appears only in females). Therefore, the ordered sets of features were identified as follows.

• MDA_Male = {Age, Temperature, BMI, RR, Pulse};
• MDA_Female = {Age, Temperature, BMI, Pulse, RR};
• MDI_Male = {Age, BMI, Temperature, Pulse, Occupation, RR};
• MDI_Female = {BMI, Age, Pulse, Temperature, RR, Occupation, Status}.

To investigate the impact of feature inclusion on model performance, features were added to the RF models based on predefined importance rankings. At each step, the model was evaluated after adding one additional feature from the previously defined set, starting with the set with important features. This cumulative inclusion strategy enabled the assessment of how each successive feature contributed to classification performance.

3.2. Model Performance

By implementing RF models across male and female groups, the features sets were first selected by LASSO, Ridge, and Elastic Net. These features were then ranked using MDA and MDI scores, and the ranked sets were subsequently used to train the models. Aspresented in

Table 3. RF model performance by cumulative feature sets in males.

MDA	Accuracy	Precision	Recall	F1-Score	Specificity	AUC
1	0.4923	0.5053	0.5057	0.4755	0.7601	0.6789
2	0.6111	0.5913	0.6163	0.5954	0.8057	0.7755
3	0.6147	0.5845	0.6071	0.5899	0.8003	0.7853
4	0.6247	0.6013	0.6249	0.6079	0.8083	0.7966
5	** 0.6600	** 0.6351	** 0.6575	** 0.6422	** 0.8262	** 0.8124

and

Table 4. RF model performance by cumulative feature sets in females.

MDA	Accuracy	Precision	Recall	F1-Score	Specificity	AUC
1	0.4617	0.4085	0.4597	0.3810	0.7267	0.6508
2	0.6333	0.5090	0.5808	0.5269	0.7744	0.7446
3	** 0.7335	** 0.5926	0.5821	0.5825	0.7779	0.7897
4	0.7236	0.5918	0.6333	** 0.6013	** 0.8036	0.7908
5	0.7153	0.5870	** 0.6384	0.5997	0.8036	** 0.8075

, which report results based on MDA, the classification performance of RF models in the test set improved as more features were included, particularly in males. Although not shown, the models using MDI-ranked features exhibited a similar pattern. The notation ** marks the best-performing models for each evaluation metric, highlighting improvements in accuracy, precision, recall, F1-score, specificity, and AUC.

Table 3 shows that the male group consistently showed higher accuracy, recall, and specificity than the female group, while females displayed slightly lower precision (Table 4). F1-score showed a similar pattern, improving with the addition of features in both groups. The AUC values remained robust across all scenarios, indicating consistent classification performance. Specificity was notably high in all models, reflecting strong negative class prediction. The results underscored the importance of feature selection. Combinations of features enhanced model robustness and predictive reliability. For example, under MDA with five features, {Age, Temperature, BMI, RR, Pulse}, the F1-score reached 0.6422 in males and 0.5997 in females. The AUC was 0.8124 for males and 0.8075 for females, showing better performance of the male model. Similarly, under MDI with six features, {BMI, Age, Pulse, temperature, RR, Occupation}, the male F1-score was 0.6293 and the female F1-score was 0.6185, with corresponding AUC values of 0.8186 and 0.8067. The male models generally performed better with a smaller difference in AUC in MDI, reflecting improved female model performance. Interestingly, female performance patterns differed from males: whereas male models improved with the inclusion of more features, female models achieved their highest values for some metrics (e.g., accuracy and precision) using only three features. This indicates that additional features did not consistently improve female model performance and, in some cases, slightly reduced it. In other words, identifying the most crucial features in predictive modeling can be more effective than simply adding more features. Notably, BMI and Pulse appeared in top-performing MDA and MDI sets, suggesting their consistent contribution to model discrimination. The sustained high specificity in both groups proved the reliability of these models in identifying non-disease cases. However, differences in F1-score and AUC revealed gender-based variations in classification effectiveness.

3.3. SHAP for Feature Importance

To interpret the experimental results obtained from the best-performing classifier, SHAP analysis results identified the most influential features driving the model’s predictions across the training and test datasets, providing explainable insights into class-level predictions. The SHAP analysis results for the male group highlighted age, RR, and temperature as the most influential features for the model’s predictions (Figure 2). Age showed the highest mean SHAP value, significantly impacting classification, followed by RR and temperature. The class-based breakdown suggested that feature importance varied by category, with pulse and BMI having lesser but notable contributions. Higher feature values were correlated with greater SHAP values, while RR, age, and BMI showed strong influence. Pulse contributed the least. For the female group, the SHAP analysis result showed that BMI, RR, and temperature were influential features (Figure 3). BMI showed the highest mean SHAP value, followed by RR and temperature, indicating their significant impact on classification. The feature set of pulse and age also contributed notably, while status and occupation had lower but observable effects across different classes.

4. Discussion

The capability of the developed model in the classification of the three major groups of diseases based on ICD-10 codes was assessed by feature selection, feature importance, and SHAP. Regarding the gender-specific risk factors, seven features were identified in the dataset.

4.1. Features and Classification

Feature selection plays a crucial role in improving classification performance. The inclusion of age, temperature, BMI, RR, and pulse from MDA and BMI, age, pulse, and temperature from MDI demonstrated their importance in distinguishing disease cases. Higher F1-scores and AUC values in male models indicated more robust classification, possibly due to feature variations. The specificity was high, indicating strong negative class detection. While male models exhibited better recall, female models indicated the necessity of feature selection to optimize disease classification based on gender-specific patterns.

4.2. Model Interpretation

Understanding model interpretation is essential for assessing the influence of different features on disease classification. The SHAP analysis results presented key predictors and showed that age, BMI, RR, and body temperature were the most influential factors. Age and RR showed stronger effects in males, while BMI and pulse played critical roles in female classifications. MDA and MDI scores confirmed these findings, reinforcing their significance in predictive modeling. The results supported model interpretability, ensuring clinically relevant and explainable predictions.

4.3. Implementation in Transitional Healthcare

The integration of machine learning-based classification models in transitional healthcare settings enhances early disease detection and patient management. By leveraging ICD-10 classifications and feature selection, hospitals can implement automated screening systems to prioritize high-risk cases [2,3,14]. Key features such as BMI, age, and pulse aid in predictive modeling for infectious diseases. The observed gender-specific variations highlighted the need for personalized healthcare strategies. These models enable timely interventions to reduce diagnostic delays and improve clinical decision-making. It is necessary to integrate real-time data to enhance the adaptability of ML models in diverse healthcare environments.

5. Conclusions

We developed a machine learning-based disease classification framework using ICD-10 codes, with an emphasis on gender-specific risk factors. Feature selection was conducted using LASSO, Ridge, and Elastic Net, followed by MDA and MDI scoring to determine feature importance. The RF classifier was trained with the training and testing dataset split into 80:20 and five-fold cross-validation, incorporating class weights to address data imbalance. Age, BMI, RR, and body temperature were key predictors for both male and female groups, while pulse had a stronger impact on female models. SHAP analysis results confirmed the differential importance of these features. Models trained with appropriate feature selection exhibited improved accuracy, recall, and F1-score, presenting the importance of balancing predictive power and interpretability. The inclusion of gender-specific variations enhances the model’s applicability to personalized medicine, supporting data-driven healthcare decision-making. It is still required to explore real-time implementations and model integration into clinical workflows. By refining predictive modeling for infectious disease classification, the developed model in this study improves early diagnosis, optimizes resource allocation, and contributes to better patient outcomes in clinical settings.