⚛️ quantum physics

Classical and Quantum Machine Learning for Population-Level Prediction of Heat-Related Physiological Events

This paper proposes a unified framework comparing classical and quantum machine learning models for predicting heat-related physiological events at the population level, finding that while classical models currently outperform due to data sparsity and imbalance, quantum models demonstrate promising non-trivial learning capabilities for future hybrid health applications.

Original authors: Saul Gonzalez-Bermejo, Tommaso Albrigi, Borja Vazquez-Morado, Urko Regueiro-Ramos, Daniel Casado-Faulı, Sergi Consul-Pacareu, Parfait Atchade-Adelomou

Published 2026-04-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Saul Gonzalez-Bermejo, Tommaso Albrigi, Borja Vazquez-Morado, Urko Regueiro-Ramos, Daniel Casado-Faulı, Sergi Consul-Pacareu, Parfait Atchade-Adelomou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict when a crowd of people will start feeling sick from the heat. It's not just about checking the thermometer; it's about understanding who is in the crowd (are they elderly? do they work outside?), where they live, and how the weather is behaving. This is a massive, messy puzzle.

This paper is essentially a race between two different types of "super-solvers" trying to solve that puzzle:

The Veteran Solver (Classical Machine Learning): This is the standard, powerful computer brain we use today. It's like a seasoned detective who has seen thousands of cases and knows exactly how to connect the dots.
The Newcomer Solver (Quantum Machine Learning): This is a futuristic, experimental brain that uses the strange laws of quantum physics. It's like a brilliant but untested apprentice who has incredible potential but is currently working with a broken flashlight and a shaky map.

Here is the story of their race, broken down simply:

1. The Challenge: A Foggy, Sparse Map

The researchers wanted to predict "heat-related physiological events" (like heat strokes or hospital visits) for entire populations.

The Problem: The data is like a foggy map. Sometimes there are no records at all (sparsity), sometimes there are huge spikes in summer (seasonality), and the data comes from different places (hospitals, weather stations, census data) that don't always fit together.
The Goal: Build a unified system that takes all these messy pieces and predicts the future accurately.

2. The Setup: Giving Them the Same Tools

To make it a fair race, the researchers didn't let the two solvers use different maps. They built a single, unified dataset combining weather, population demographics (age, gender), and economic data (what jobs people have).

They cleaned the data, removed duplicates, and compressed it (like zipping a huge file into a smaller one) so both solvers could read it easily.
They trained both solvers on data from the USA (where there was lots of data) and then asked them to predict what would happen in Catalonia, Spain (where data was scarcer).

3. The Race: Who Wins?

The Veteran (Classical Machine Learning)

The classical model (specifically a "LightGBM" algorithm) is like a master chef. It knows exactly how to mix ingredients.

Performance: It did a very good job. It made predictions that were close to reality.
Why? It's great at handling "tabular data" (rows and columns of numbers) and can spot complex patterns in messy data without getting confused.

The Newcomer (Quantum Machine Learning)

The quantum model uses "variational circuits." Imagine this as a machine that tries to learn by spinning coins in a quantum superposition (where they are heads and tails at the same time) to find the best answer.

Performance: It did okay, but it wasn't as accurate as the veteran. Its predictions were more scattered, like a dart thrower who hits the board but misses the bullseye.
The Catch: The quantum computer they used is currently in the "NISQ" era (Noisy Intermediate-Scale Quantum). Think of this as trying to build a skyscraper with a wobbly ladder and a hammer that sometimes misses. The hardware is still too small and too "noisy" to handle the full complexity of the problem perfectly.

4. The Verdict: A Reality Check

The paper concludes with a very honest message: The Veteran wins right now.

Classical models are still the kings of this specific task. They are more accurate, especially when the data is sparse or the events are rare.
Quantum models didn't fail completely. They showed they can learn and find patterns. They aren't just guessing randomly; they are capturing some of the "shape" of the problem. However, they are currently limited by the fact that quantum computers are still in their "infancy."

The Big Picture Analogy

Imagine you are trying to navigate a stormy ocean.

Classical ML is a massive, modern cruise ship with GPS, radar, and a seasoned captain. It gets you to the destination safely and on time.
Quantum ML is a sleek, experimental speedboat with a revolutionary new engine. The engine is amazing and could theoretically be faster, but right now, the boat is small, the engine sputters in the waves, and the navigation system is glitchy.

The Takeaway:
This paper isn't saying "Quantum is useless." It's saying, "Quantum is promising, but we need to wait for the hardware to mature before it can beat the classics at this specific job."

The researchers have built a playground where both technologies can run side-by-side. As quantum computers get bigger and less "noisy" (like the speedboat getting a better engine), this playground will be ready to see if the newcomer can finally overtake the veteran. For now, if you need to predict heat illness in a crowd, you should trust the veteran.

1. Problem Statement

The paper addresses the challenge of predicting heat-related physiological events (HRIs) at the population level. This task is statistically difficult due to:

Complex Interactions: The interplay between climatic, demographic, and socioeconomic factors.
Data Sparsity and Seasonality: Observational data is often sparse, aggregated at mismatched geographic scales, and heavily concentrated during summer months.
Zero-Inflation: The target variable (HRI counts) contains a high proportion of zeros and low absolute counts, making standard regression models prone to conservative predictions.
Heterogeneity: Integrating diverse data sources (meteorological, census, labor, and hospitalization records) into a unified predictive framework.

The core research question is not to claim "quantum advantage," but to empirically evaluate to what extent current Quantum Machine Learning (QML) architectures can approximate or complement classical performance in this specific, realistic public health domain under Noisy Intermediate-Scale Quantum (NISQ) constraints.

2. Methodology

The authors propose a unified predictive framework where both classical and quantum models share an identical data pipeline to ensure a fair comparison.

A. Data Construction & Feature Engineering

The study constructs a unified weekly county-level dataset integrating:

Climatic Variables: Max/mean/min temperature, vapor pressure, humidity, heatwave indicators, and days exceeding the 95th percentile.
Demographic Variables: Total population and age/gender breakdowns (converted to ratios to mitigate scale effects).
Socioeconomic Variables: Employment sector distributions (agriculture, construction, industry, services).
Synthetic Seasonal Features: To address data sparsity, the authors modeled HRI intensity using a Gaussian seasonal kernel (representing the annual cycle) and a heatwave response kernel (exponential decay). This provides the model with structured prior knowledge about seasonality.

B. Common Preprocessing Pipeline

To ensure compatibility with both paradigms:

Standardization: Z-score normalization is applied to all features. This is critical for QML, as features are encoded as rotation angles.
Correlation Analysis: Redundant variables are filtered to reduce multicollinearity.
Dimensionality Reduction (PCA): Principal Component Analysis is used to compress the feature space. The number of components ( $k$ ) is selected to retain 98% of the variance. This step is crucial for QML to fit within the limited qubit counts of current hardware.

C. Model Architectures

Classical Baseline: A LightGBM (Light Gradient Boosting Machine) model. This was chosen as the benchmark due to its proven efficacy on structured tabular data with non-linear dependencies.
Quantum Model: A Variational Quantum Sequential Model (QSM).
- Encoding: Uses Angle Embedding to map PCA-reduced classical features into quantum states via $R_Y$ rotations.
- Architecture: Composed of parameterized variational layers containing $R_X, R_Y, R_Z$ rotations and entangling gates (CNOT).
- Data Re-uploading: The input features are injected multiple times throughout the circuit layers (rather than just once at the start) to increase expressivity without adding qubits.
- Measurement: Predictions are derived from the expectation values of observables, mapped to a final regression output.
- Training: Both models are trained to minimize Mean Squared Error (MSE).

D. Experimental Setup

Training Data: A large dataset from the United States (CDC health data, Daymet climate, US Census).
Testing/Inference Data: A dataset from Catalonia, Spain (Meteocat, Idescat, Hospital Morbidity Survey).
Strategy: Models are trained on US data (filtered to match Catalonia's climate) and evaluated on the Catalonia dataset to test generalization across regions.

3. Key Contributions

Unified Pipeline: Creation of a reproducible pipeline that integrates heterogeneous public health and climate data, specifically designed to support both classical and quantum learning.
Fair Comparison: Implementation of a shared preprocessing strategy (standardization + PCA) ensuring that both models operate on the exact same feature space.
Empirical Benchmarking: A rigorous comparative study evaluating QML against a strong classical baseline (LightGBM) on a real-world, sparse, and zero-inflated health prediction task.
Methodological Foundation: Establishment of a framework for future hybrid health modeling as quantum hardware evolves.

4. Results

The study evaluated performance using Mean Absolute Error (MAE) and the Coefficient of Determination ( $R^2$ ).

Catalonia Dataset (Sparse/Small Scale):
- Classical (LightGBM): MAE = 0.33
- Quantum (QSM): MAE = 2.84
- Observation: Both models struggled (negative $R^2$ ), but the classical model was significantly more stable and accurate.
USA Dataset (Larger Scale):
- Classical (LightGBM): MAE = 0.54
- Quantum (QSM): MAE = 2.57
- Observation: The classical model demonstrated clear superiority in explanatory power and stability. The quantum model showed higher error dispersion.

Key Finding: While the quantum model did not outperform the classical baseline, it demonstrated non-trivial learning capability, capturing meaningful predictive structure rather than behaving randomly. The performance gap is attributed to NISQ hardware limitations (noise, limited depth, qubit count) and the extreme sparsity of the target variable.

5. Significance and Conclusion

Realistic Assessment: The paper avoids hype, providing a grounded assessment that current QML is not yet ready to replace classical methods for complex, sparse population health tasks.
NISQ Constraints: It highlights that while variational circuits with data re-uploading are theoretically expressive, they are currently bottlenecked by hardware noise and circuit depth limitations.
Future Outlook: The work establishes a methodological baseline for hybrid quantum-classical models. As quantum hardware improves (reduced noise, more qubits), these pipelines may become viable for problems where classical models struggle with high-dimensional, non-linear feature interactions.
Public Health Impact: The framework offers a robust tool for territorial risk assessment and resource planning, even if the current quantum component serves primarily as an experimental benchmark.

In summary, the paper successfully demonstrates that while classical machine learning (specifically boosting methods) currently dominates in population-level heat prediction, quantum machine learning is capable of learning non-trivial patterns, laying the groundwork for future hybrid approaches as quantum technology matures.