The assessment of patient health status through physiological parameters like Oxygen Saturation (SpO₂) and Heart Rate (HR) remains a cornerstone of clinical practice. Traditionally, this process relies on manual interpretation by healthcare professionals, which despite its widespread use suffers from inherent limitations in objectivity, reproducibility and time efficiency. Subjective variability among clinicians and the labor-intensive nature of continuous monitoring often lead to delayed interventions, particularly in high-volume clinical settings.

In recent years, Machine Learning (ML) has emerged as a transformative tool for automating medical diagnostics, offering the potential to overcome these challenges while maintaining (or even improving) diagnostic accuracy [1,2]. Supervised learning algorithms, for instance, have demonstrated remarkable success in classifying conditions such as hypoxia, bradycardia and tachycardia from physiological data [3]. However, a critical barrier persists: many state-of-the-art ML models, particularly "black-box" systems like deep neural networks, lack interpretability a non-negotiable requirement for clinical adoption [4].

Clinicians must understand why a model makes specific decisions to trust and act upon its outputs, especially in life-critical scenarios.

For this study, we selected a real-world dataset comprising a population of 1,000 individuals displaying the relevant clinical symptoms under investigation [6].

This study addresses this gap by proposing a transparent, decision-tree-based classification system for health status assessment using SpO₂ and HR. Our approach uniquely balances high performance with interpretability, providing clinicians with actionable insights through clear decision rules ("SpO₂≤92.5% → Hypoxia") [5]. The specific objectives.

The specific objectives of this study are:

• To develop an interpretable classification model based on SpO₂ and HR parameters
• To rigorously evaluate its diagnostic performance
• To identify optimal clinical thresholds for each health status
• To propose pathways for integration into clinical workflows

Data Preparation

The raw physiological dataset underwent a rigorous preprocessing pipeline to ensure robustness and mitigate biases inherent in medical data. This critical phase comprised four methodical stages [6].

Quality Control and Artifact Removal

A two-step filtration process was implemented:

• Automated rejection of signal segments with physiologically implausible values (HR<30 bpm or SpO₂ < 70%) using threshold-based detection
• Manual verification by clinical experts to eliminate motion artifacts and sensor malfunctions, preserving only data segments with ≥95% signal integrity

Z-Score Normalization

Each parameter was standardized using:

where, μ and σ represent the population-wise mean and standard deviation, respectively.

This transformation addressed scale variance between HR (60–100 bpm) and SpO₂ (0–100%) while maintaining clinical interpretability of thresholds.

Clinical Label Encoding

A hierarchical labeling system was developed in collaboration with pulmonologists and cardiologists:

• Level 1: Binary classification (normal/abnormal)
• Level 2: Condition-specific categorization (hypoxia, bradycardia, tachycardia)
• Level 3: Severity grading (mild/moderate/severe) based on clinical guidelines

Stratified Splitting (70% Training / 30% Testing)

The dataset was partitioned using scikit-learns Stratified Shuffle Split to:

• Preserve the original distribution of all diagnostic categories in both subsets
• Prevent leakage between training and testing cohorts
• Ensure ≥100 samples per condition in the test set for statistical power

Rationale

This preprocessing sequence optimizes model generalizability while adhering to Findable, Accessible, Interoperable, Reusable (FAIR) data principles for medical AI [5].

Exploratory Analysis

The Figure 1 show the distributions of frequency of SpO₂ and HR (Table 1).

The relationship between SpO₂ and HR by health status are showing on the Figure 2.

In Table 4, we will present a classification report, as well as the resulting confusion matrix.

Confusion Matrix:

Decision Tree Model Architecture

The classification framework employed a Classification and Regression Tree (CART) algorithm, selected for its dual capability to handle both categorical and continuous variables while generating human-interpretable decision rules [7,8]. The model architecture was optimized through systematic hyperparameter tuning and validation protocols:

• Core Algorithm Specifications
• Splitting Criterion
• Utilized the Gini impurity index (Gini = 1 - Σ(pi²)) as the primary splitting metric

Advantages over entropy-based approaches:

• Computationally efficient (no logarithmic calculations)
• Strong bias toward balanced splits in medical data distributions

Depth Control

Constrained to maximum depth of 4 to:

• Prevent overfitting to training set noise
• Ensure clinical usability (≤5 sequential decisions align with clinician cognitive workflows)
• Maintain compliance with the "15-second rule" for bedside interpretation

Node Splitting Policy

Minimum samples per split = 5 to:

• Preserve statistical significance (α<0.05 for x² tests of parameter thresholds)
• Exclude clinically irrelevant subgroups (n<5% of cohort) Validation Protocol

Five fold stratified cross-validation with:

• Repeated random subsampling (3 iterations)
• Preservation of class proportions in each fold
• Early stopping if validation accuracy plateaued (±0.5% over 10 epochs) Mathematical Optimization

Table 1: Data Overview

Figure 1: Distributions of frequency of SPO₂ and HR

Figure 2: Relationship between SPO₂ and HR by health status

Table 2: Descriptive Statistics

Table 3: Class Distribution

The tree construction solved the following minimization problem at each node t:

where:

s          =    Optimal splitting threshold

t_O, t_O   =    Left/right child nodes

n_O, n_O =    Sample counts in child nodes

N         =    Parent node sample size

Clinical Interpretability Enhancements:

• Post-pruning of branches with <2% population coverage
• Conversion of Z-score thresholds back to physiological units ("SpO₂≤92.5%") for clinical deployment
• Integration of SHAP values to quantify feature importance while maintaining tree structure

Comparative Advantage

This configuration achieved 96.2% accuracy while requiring only 7 binary decision rules - significantly more interpretable than equivalent random forest (23 rules) or neural network (opaque) models (Figure 3). We implemented a Classification and Regression Tree (CART) classifier with:

• Splitting criterion: Gini index
• Maximum depth: 4
• Minimum samples to split a node: 5
• 5-fold cross-validation for optimization

Clinical Interpretation

Interpretation of Results

Key Decision Rules

:

|--- SpO₂ <= 95.00

|        |--- HR <= 60.00

|        |        |--- class: 3

|        |--- HR >      60.00

|        |        |--- SpO₂ <= 89.96

|        |        |        |--- class: 3

|        |        |--- SpO₂ >         89.96

|        |        |        |--- HR <= 100.05

|        |        |        |        |--- class: 2

|        |        |        |--- HR >      100.05

|        |        |        |        |--- class: 4

|--- SpO₂ >         95.00

|        |--- HR <= 59.93

|        |        |--- class: 0

|           |--- HR >      59.93

|        |        |--- HR <= 100.15

|        |        |        |--- class: 1

|        |        |--- HR >      100.15

|             |             |             |--- class: 4

Feature Importance

• SpO₂: 0.47
• HR: 0.53

Clinical Recommendations:

:

- SpO₂ < 92%: hypoxia

- Une HR < 60 bpm: bradycardia

- Une HR > 100 bpm: tachycardia

- Healthy cases typically have SpO₂> 95% and HR between 60-100 bpm

Limitations and Future Directions

Main Limitations Include

Need for validation on independent cohorts-Lack of comorbidity consideration [9,10]. Generalizability to pediatric populations Future research could integrate.

Additional parameters (blood pressure, temperature)-Hybrid ap proaches combining decision trees and neural networks (Figure 4).

This study demonstrates that a rigorously optimized decision tree model can achieve high- performance automated classification of patient health status using only two core physiological parameters oxygen saturation (SpO₂) and Heart Rate (HR). With an accuracy of 96.2% and clinically interpretable decision rules (SpO₂ ≤ 92.5% → Hypoxia), our framework addresses a critical need in medical artificial intelligence: bridging the gap between algorithmic performance and clinical usability.

Table 4: Classification Report

Figure 3: Decision tree

Figure 4: Decision boundaries

Key Advancements

Interpretability-Preserving Performance

• The CART-derived model matches the diagnostic accuracy of "black-box" alternatives (neural networks) while generating actionable decision thresholds that align with existing clinical guidelines
• Each classification pathway requires ≤4 sequential decisions-compatible with real-time bedside triage workflows

Clinical Integration Potential

Demonstrated feasibility for deployment in:

• Telemedicine platforms (low computational overhead)
• ICU dashboards (rules map directly to alarm thresholds)
• EMR decision support (interpretability meets regulatory requirements)

Limitations and Future Directions

While promising, these findings warrant further validation through:

• Multicenter clinical trials assessing:
• Impact on clinician decision latency (target: ≥30% reduction)
• Effect on patient outcomes (e.g., early intervention rates)
• Expansion to comorbid populations (current study excluded patients with concurrent cardiovascular/respiratory conditions)
• Integration of additional parameters (e.g., blood pressure, etCO₂) to enhance specificity in borderline cases

Broader Implications

This work provides a proof-of-concept framework for developing interpretable-by-design medical AI systems. Future iterations could adopt hybrid architectures (e.g., decision trees guiding neural network attention) to balance accuracy and transparency in more complex diagnostic scenarios.

1. Alghamdi, M., et al. "Predicting cardiovascular diseases using decision tree-based machine learning models." Journal of Medical Systems, vol. 41, no. 2, 2017, p. 30.
2. Ruban, V. and Krithi K. "Heart disease prediction using machine learning models." International Journal of Recent Technology and Engineering (IJRTE), vol. 8, no. 5S, 2020, pp. 2277–3878.
3. Islam, M.M. et al. "Development of smart healthcare monitoring system in IoT environment." SN Computer Science, vol. 1, no. 3, 2020, p. 185.
4. Breiman, L., et al. Classification and regression trees. Chapman & Hall, 1984.
5. Kumar, P.M., et al. "Intelligent face recognition and navigation system using neural learning for smart security in Internet of Things." Cluster Computing, vol. 22, suppl. 1, 2018, pp. 7733–7744.
6. Pattekari, S.A. and A. Parveen. "Prediction system for heart disease using naïve Bayes and decision tree." International Journal of Advanced Computer Research, vol. 2, no. 3, 2012, pp. 31–35.
7. Saxena, K. and Sharma R. "Efficient heart disease prediction system using decision tree." Proceedings of the International Conference on Computing, Communication & Automation (ICCCA), 2016, pp. 72–77.
8. Loh, W. Y. "Classification and regression trees." WIRES Data Mining and Knowledge Discovery, vol. 1, no. 1, 2011, pp. 14–23.
9. Rajpurkar, P., et al. "AI in health and medicine." Nature Medicine, vol. 28, no. 1, 2022, pp. 31–38.
10. Mahdab, S. and Moualdia A. "Gestion intelligente d’un système hybride par la logique floue: Application au soudage à l’arc." Revue Roumaine des Sciences Techniques – Électrotechnique et Énergétique, vol. 67, no. 2, 2022, pp. 111–116.