High-Performance Classification of Mpox Symptoms Using Support Vector Classifier and Quadratic Discriminant Analysis

This study demonstrates that Support Vector Classifier and Quadratic Discriminant Analysis models, trained on clinical symptom data, achieve high accuracy (97.7%) and robust discriminatory power in detecting Mpox, with skin rash identified as the most predictive feature.

The Gist

Imagine you are a doctor in a remote village. A patient walks in with a fever and a rash. Is it just a bad case of chickenpox? Is it measles? Or is it the dangerous Mpox virus?

In the real world, the only way to know for sure is to send a sample to a high-tech lab for a DNA test (PCR). But imagine if that lab is 500 miles away, costs a fortune, or simply doesn't exist. You're left guessing, and in a fast-moving outbreak, a wrong guess can be deadly.

This paper is like a team of digital detectives trying to solve that guessing game using Machine Learning. Here is the story of how they did it, explained simply.

🕵️‍♂️ The Mission: Building a "Symptom Detective"

The researchers wanted to build a computer program that could look at a patient's list of symptoms and say, "99% chance this is Mpox," or "This is likely something else," without needing a lab test.

They treated the computer models like trainees at a police academy. They gave these trainees a massive stack of case files (data from the World Health Organization) containing information on hundreds of people: their age, gender, country, and a long list of symptoms (fever, headache, rash, etc.).

🧹 The Cleanup: Organizing the Messy Evidence

Real-world data is messy. One file might say "muscle pain," another "myalgia," and a third "body ache." It's like trying to solve a puzzle where some pieces are labeled in three different languages.

The researchers acted as librarians:

1. Standardized the language: They made sure "muscle pain" and "myalgia" were treated as the same thing.
2. Fixed the gaps: If a file was missing a patient's age, they used the average age of similar patients to fill it in.
3. Balanced the scales: They made sure they had an equal number of "confirmed Mpox" cases and "suspected but not Mpox" cases, so the computer didn't get biased toward one side.

🏆 The Competition: Who is the Best Detective?

Once the data was clean, they introduced five different types of "detective algorithms" to the case:

1. Decision Tree: A flowchart that asks "Yes/No" questions (e.g., "Do you have a rash?").
2. Extra Trees: A team of many decision trees voting on the answer.
3. Perceptron: A simple neural network that learns by trial and error.
4. Quadratic Discriminant Analysis (QDA): A math-heavy detective that looks for complex patterns and curves in the data.
5. Support Vector Classifier (SVC): A detective that tries to draw the perfect line in the sand to separate "Mpox" from "Not Mpox."

They ran a race to see which one could identify the virus most accurately.

🏅 The Winners: The "Perfect Score" Trio

The results were surprising and exciting. Three of the detectives tied for first place, achieving a 97.7% accuracy rate.

• The Winners: Support Vector Classifier (SVC), Quadratic Discriminant Analysis (QDA), and Perceptron.
• The Scorecard: They got almost every single case right. Out of the test cases, they correctly identified 44 people with Mpox and didn't mistakenly accuse a single healthy person of having it (zero false alarms). They only missed 2 real cases (false negatives), which is incredibly low.

Think of it like a security scanner at an airport. If you have 100 people with a weapon, this scanner catches 98 of them and never falsely alarms on a harmless toothbrush.

🔍 The "Aha!" Moment: What Matters Most?

The researchers asked the winning computers: "What clue made you so sure?"

The answer was clear: The Skin Rash.
Just like a detective knowing that a muddy footprint is the most important clue at a crime scene, the computer learned that skin rash was the single most important symptom (with a score of 0.12). It was followed closely by skin lesions and fever.

Interestingly, swollen lymph nodes (a classic Mpox sign) was ranked much lower. The researchers suspect this isn't because it's unimportant, but because people in the data often forgot to report it, or doctors didn't write it down. It's a reminder that even the smartest AI is only as good as the notes it reads.

🌍 What This Means for the Real World

The study also found some interesting patterns about who is getting sick:

• Age: Most cases were in adults aged 20–64.
• Gender: More men were affected than women.
• Location: The Democratic Republic of Congo had the highest number of cases in their data.

🚀 The Big Picture

This paper isn't saying we should stop using lab tests. Instead, it's proposing a super-powered triage tool.

Imagine a health worker in a remote village with a tablet. They type in the patient's symptoms. The app (powered by these winning algorithms) instantly says, "High probability of Mpox." This tells the health worker, "Don't wait for the lab; isolate this patient immediately and send them to the nearest facility."

In short: By teaching computers to read the "story" told by symptoms, we can catch Mpox faster, cheaper, and in places where high-tech labs can't reach. It's like giving every village doctor a super-detective in their pocket.

Technical Summary

1. Problem Statement

The global resurgence of Mpox (Monkeypox), particularly the emergence of the highly virulent Clade I (specifically subclade IB), has created significant diagnostic challenges.

• Diagnostic Bottlenecks: Conventional molecular diagnostics (PCR) are costly, labor-intensive, and often inaccessible in resource-limited settings (e.g., the Democratic Republic of Congo), where only a small fraction of suspected cases receive laboratory confirmation.
• Clinical Ambiguity: Mpox symptoms (fever, rash, lymphadenopathy) closely mimic other exanthematous diseases like chickenpox, measles, and bacterial skin infections, leading to high rates of misdiagnosis (up to 50% in endemic regions).
• Need for Alternatives: There is an urgent need for scalable, cost-effective, and non-invasive diagnostic tools that rely on clinical symptomatology rather than expensive molecular testing to enhance early detection and surveillance.

2. Methodology

The study employed a supervised machine learning (ML) approach to classify Mpox cases based on clinical symptoms and demographic data.

• Data Source & Collection:

  • Dataset: An open-access, anonymized line-listed dataset of suspected and confirmed Mpox cases (2022–2024) obtained from the WHO via the Global.Health platform.
  • Preprocessing:
    • Cleaning: Removed extraneous variables; standardized heterogeneous symptom terminology (e.g., merging "muscle pain," "muscle ache," and "myalgia").
    • Encoding: Symptom variables were binarized (Yes=1, No=0). The target variable 'Status' was encoded (Confirmed=1, Suspected=0).
    • Feature Engineering: Age was grouped into three brackets (0–19, 20–64, ≥65). Missing values in 'Gender' and 'Age' were imputed using the mode.
    • Balancing: To mitigate class imbalance, suspected cases were randomly downsampled to match the number of confirmed cases, resulting in a balanced dataset of 438 individuals (219 confirmed, 219 suspected) with 48 features (42 symptoms + 6 demographics).
  • Splitting: The data was partitioned into training (80%), validation (10%), and testing (10%) sets. Z-score normalization was applied to all features.

• Model Selection & Training:

  • Initial Screening: 26 classifiers were evaluated using the LazyPredict library.
  • Top Models: Five models were selected for rigorous evaluation: Extra Trees (ETC), Quadratic Discriminant Analysis (QDA), Decision Tree (DTC), Perceptron, and Support Vector Classifier (SVC).
  • Optimization: Hyperparameter tuning was performed using GridSearchCV with cross-validation to prevent overfitting.

• Evaluation Metrics:

  • Performance was assessed using Accuracy, F1-score, Recall, and Area Under the Receiver Operating Characteristic Curve (ROC-AUC).
  • Feature Importance: Permutation-based analysis was used to determine the contribution of specific clinical symptoms to model predictions.

3. Key Results

• Model Performance:

  • Three models achieved superior and identical performance metrics: SVC, QDA, and Perceptron.
  • Metrics: All three achieved 97.7% Accuracy, 97.7% F1-score, 97.7% ROC-AUC, and 95.5% Recall.
  • Confusion Matrix Analysis: These top models correctly identified 44 True Positives with zero False Positives. They produced only 2 False Negatives and 42 True Negatives, demonstrating exceptional specificity and sensitivity.
  • Comparison: The Decision Tree (94.3% accuracy) and Extra Trees (95.5% accuracy) performed well but were outperformed by the top three.

• Feature Importance:

  • Skin Rash was identified as the most predictive feature (Permutation Importance score: 0.12).
  • Skin Lesions and Fever followed closely (score: 0.11 each).
  • Surprising Finding: Lymphadenopathy, a hallmark clinical sign distinguishing Mpox from other diseases, ranked only 11th (score < 0.02), suggesting it may be inconsistently reported in the source data or less discriminatory in this specific dataset context.

• Epidemiological Insights:

  • Demographics: The dataset showed a higher prevalence in males and the 20–64 age group, contrasting with historical African surveillance data which often highlights children. This may reflect reporting biases or the influence of sexual transmission networks in recent outbreaks.
  • Geography: The Democratic Republic of Congo (DRC) had the highest case load, followed by Canada, Nigeria, and Portugal.

4. Key Contributions

1. High-Performance Clinical ML: Demonstrated that simple clinical symptom data, without genomic or imaging inputs, can achieve >97% accuracy in distinguishing Mpox from suspected non-Mpox cases using standard ML algorithms (SVC, QDA).
2. Algorithmic Benchmarking: Provided a comparative analysis of five distinct classifiers, identifying QDA and SVC as the most robust for this specific diagnostic task.
3. Feature Analysis: Quantified the predictive power of specific symptoms, validating the centrality of dermatological symptoms (rash, lesions) while highlighting potential data quality issues regarding lymphadenopathy reporting.
4. Scalable Diagnostic Framework: Proposed a low-cost, non-invasive diagnostic workflow suitable for resource-constrained settings where PCR is unavailable.

5. Significance and Limitations

• Significance:

  • Public Health Impact: These models could serve as rapid triage tools in primary care settings, enabling earlier isolation and containment of outbreaks, particularly in regions with limited laboratory infrastructure.
  • Clinical Decision Support: The high recall (95.5%) ensures that true cases are rarely missed, a critical requirement for infectious disease control.
  • Future Integration: The study suggests that integrating these symptom-based models with genomic or image-based tools could create a multi-modal diagnostic system.

• Limitations:

  • Data Quality: The study relied on retrospective, self-reported, or heterogeneous clinical data, introducing potential recall bias and missing values.
  • Generalizability: The dataset may not fully represent the global burden due to uneven reporting practices. The demographic skew (adults > children) may limit applicability to pediatric-focused endemic regions without further validation.
  • Causality: As an observational study, it cannot establish causal relationships.
  • Validation: The models require prospective validation in real-world clinical environments before deployment.

Conclusion: The study successfully validates the potential of Support Vector Classifiers and Quadratic Discriminant Analysis to act as high-accuracy, cost-effective screening tools for Mpox based solely on clinical symptoms, offering a viable pathway to improve disease surveillance in resource-limited settings.