Lesson 8.1.2: Estimating disease intensity from hyperspectral features

Definition of Disease Intensity

In plant pathology, disease intensity refers to the severity of infection expressed as a quantitative measure of the proportion of plant tissue that is affected by a particular pathogen. Unlike disease incidence, which simply records the presence or absence of disease on a plant or plot, intensity captures how extensive the damage is. This distinction is vital for crop management because it informs both the urgency of intervention and the choice of control strategy.

Why Disease Intensity Matters for Crop Management

Effective crop management relies on accurate, timely information about disease pressure. Disease intensity provides several key pieces of information:

• Resource allocation – High intensity levels indicate that a significant portion of the crop is compromised, signalling that fungicide application or other control measures should be prioritised in those areas to prevent yield loss.
• Threshold setting – Many decision-support tools use intensity thresholds to trigger treatment. For example, a threshold of 30 % leaf area infected may prompt a fungicide spray, whereas a lower threshold might only warrant monitoring.
• Evaluation of control measures – By comparing pre- and post-treatment intensity, growers can assess the efficacy of a fungicide or cultural practice. A reduction from 45 % to 15 % intensity demonstrates a successful intervention.
• Risk modelling – Intensity data feed into predictive models that estimate potential yield loss. Models such as the Integrated Disease Management (IDM) framework use intensity as a primary variable to forecast economic thresholds.

Quantifying Disease Intensity: Common Methods

There are several established approaches to measuring intensity, each with its own advantages and limitations. The choice of method often depends on the crop, disease, and available resources.

1. Visual Scoring Systems – The most straightforward technique involves a trained observer estimating the percentage of leaf area or plant parts affected. For instance, the Modified Cobb Scale is widely used for fungal leaf spot diseases, assigning scores from 0 % (no infection) to 100 % (complete canopy coverage). Visual scoring is inexpensive and can be performed in the field, but it is subjective and subject to inter‑observer variability.
2. Digital Image Analysis – High-resolution photographs of leaves or canopy sections are processed using specialised software to calculate the proportion of diseased pixels. This method offers high precision and repeatability. However, it requires imaging equipment, lighting control, and software licences, which may be prohibitive for smallholders.
3. Spectroscopic Methods – Hyperspectral imaging captures reflectance across many narrow wavelength bands, allowing the detection of subtle biochemical changes associated with infection. By applying machine learning algorithms, researchers can estimate disease intensity with high accuracy. The main drawbacks are the cost of hyperspectral cameras and the computational expertise required for data processing.
4. Biomass or Yield Loss Measurements – In some cases, disease intensity is inferred from the reduction in biomass or yield compared to a healthy control plot. This approach integrates the cumulative effect of disease over time but cannot isolate the intensity at a specific point in the growth cycle.

Comparing Visual and Spectral Approaches: A Case Study

Consider a wheat field affected by Septoria tritici blotch (STB). Traditional visual scoring by a field technician might yield the following results over a 10 ha plot:

In contrast, a hyperspectral survey conducted a week later produced intensity estimates based on a trained neural network model:

The close agreement between the two methods suggests that, for this disease and crop, visual scoring is sufficiently accurate for routine monitoring. However, in a scenario where the disease manifests as a subtle chlorosis that is difficult to quantify visually, the spectral method may provide earlier detection and more precise intensity estimates.

Integrating Disease Intensity into Decision-Making Frameworks

Modern precision agriculture platforms incorporate disease intensity metrics into decision support systems (DSS). A typical workflow might involve:

1. Data acquisition – Sensors (visual cameras, drones, or hyperspectral units) collect imagery across the field.
2. Data processing – Images are processed to extract intensity values, often using automated pipelines that apply segmentation, classification, and regression algorithms.
3. Threshold evaluation – The DSS compares the processed intensity against pre‑defined economic thresholds. For example, if the average intensity exceeds 35 %, the system recommends a fungicide spray.
4. Action planning – The platform generates a georeferenced map that guides variable-rate application, allowing growers to concentrate fungicide on high‑intensity zones while avoiding unnecessary treatment elsewhere.

By integrating intensity data, growers can move from a reactive approach—treating after visible damage—to a proactive, risk‑based strategy that optimises input use and reduces environmental impact.

Challenges and Considerations

While the concept of disease intensity is straightforward, its practical application is fraught with challenges:

• Temporal dynamics – Disease intensity fluctuates over time as the pathogen progresses and as the crop responds. A single measurement may not capture these dynamics, leading to misinformed decisions.
• Spatial heterogeneity – Intensity can vary dramatically within a field due to microclimate differences, plant density, or management history. High-resolution sampling is essential to avoid underestimating hotspots.
• Observer bias – In visual scoring, even experienced technicians can over‑ or under‑estimate intensity, especially in the early stages of infection. Training and standardised protocols mitigate but do not eliminate this bias.
• Cost–benefit trade‑offs – Advanced methods such as hyperspectral imaging offer precision but come at higher costs. Growers must weigh the marginal benefits against the investment required.

Practical Takeaway for Growers

For most commercial operations, a hybrid approach often delivers the best balance between accuracy and practicality:

1. Use visual scoring as a rapid, low‑cost screening tool to identify plots that may require further investigation.
2. Deploy spectral or image‑analysis methods in those high‑risk plots to refine intensity estimates and guide variable‑rate applications.
3. Continuously calibrate visual scores against spectral data to improve observer accuracy over time.

By embedding disease intensity into the crop management routine, growers can optimise input use, reduce unnecessary treatments, and ultimately protect yields from the invisible threat of plant disease.

Mechanisms of Hyperspectral Detection

Hyperspectral imaging captures reflected light across hundreds of contiguous wavelength bands, spanning the visible to short‑wave infrared spectrum. By analysing the fine structure of the reflectance spectrum, one can infer subtle changes in leaf chemistry, anatomy and physiology that are caused by disease. The key to disease detection lies in recognising how pathogens perturb the biophysical properties of foliage and how these perturbations translate into spectral signatures.

1. Chemical Alterations and Spectral Consequences

Pathogens often modify the biochemical composition of infected tissue. Two principal chemical changes are of interest: chlorophyll degradation and the accumulation of secondary metabolites such as phenolics or pathogen‑derived pigments.

1. Chlorophyll and Carotenoids – Viral or bacterial infections typically induce chlorosis, characterised by a decline in chlorophyll a and b. This reduces absorption in the 400–700 nm range, especially near 670 nm where chlorophyll‑a absorbs strongly. Hyperspectral data show a corresponding increase in reflectance in the red edge (680–750 nm) and a characteristic “red‑edge shift” toward longer wavelengths. In contrast, fungal infections that trigger necrosis may leave chlorophyll intact until the tissue dies, producing a later, more pronounced spectral change.
2. Phenolic Accumulation – Many plants synthesise phenolic compounds as a defence. These molecules absorb strongly in the UV and blue regions (<450 nm) and can also increase near‑infrared (NIR) reflectance due to changes in leaf structure. For example, in grapevine powdery mildew, a subtle rise in reflectance around 500 nm can be detected before visible symptoms appear.

By quantifying the depth of the chlorophyll absorption band and the slope of the red edge, algorithms can estimate disease severity even when visible symptoms are absent.

2. Structural Modifications and Their Spectral Footprint

Beyond chemistry, many diseases remodel leaf architecture. These structural changes influence light scattering and absorption.

• Cell Wall Integrity – Pathogens such as rust fungi often cause cell wall loosening. This increases the leaf’s light‑scattering capacity, leading to higher reflectance in the NIR (700–1300 nm). The effect can be quantified using the NIR ratio (NIR/Red) which rises with disease progression.
• Water Content – Bacterial blight may cause wilting, reducing leaf water content. Water loss reduces absorption in the 1400–1700 nm band, resulting in a flatter spectrum in the short‑wave infrared. Hyperspectral indices like the Normalised Difference Water Index (NDWI) are sensitive to such changes.
• Leaf Thickness and Porosity – Some fungal infections thicken the epidermis or create pores. These morphological alterations change the canopy’s albedo, particularly in the 800–900 nm window where the reflectance is highly sensitive to leaf thickness.

Structural changes often precede biochemical shifts, providing an early warning in hyperspectral data.

3. Comparative Sensitivity of Spectral Regions

Not all wavelengths are equally informative for disease detection. A comparison of spectral regions highlights their relative strengths.

Combining indices from multiple regions often yields the best detection performance. For instance, the Normalised Difference Red Edge (NDRE) and the NIR‑Red ratio together can differentiate between early chlorosis and advanced necrosis.

4. Case Studies Illustrating Spectral Mechanisms

Concrete examples help visualise the theory.

“In a study of wheat infected with Fusarium head blight, hyperspectral imaging revealed a distinct increase in reflectance at 680 nm and a pronounced red‑edge shift within 48 hours of inoculation. Subsequent grain analysis confirmed mycotoxin levels, demonstrating the method’s predictive power.” – Smith et al., 2021

Another example involves late blight in potatoes. The pathogen induces rapid cell death, leading to a sharp drop in chlorophyll absorption and a spike in NIR reflectance within 24 hours. Algorithms that monitor the ratio of 750 nm to 800 nm can flag affected areas before visual lesions appear.

5. Limitations and Mitigation Strategies

While hyperspectral detection is powerful, several challenges must be addressed.

1. Background Variability – Soil reflectance or surrounding vegetation can mask subtle leaf changes. Spectral unmixing techniques separate leaf spectra from background by modelling the scene as a linear combination of endmembers.
2. Atmospheric Effects – For airborne or satellite platforms, atmospheric scattering can distort the spectrum. Radiometric correction using atmospheric models (e.g., MODTRAN) restores true reflectance values.
3. Temporal Dynamics – Disease progression is rapid in some crops. High‑frequency monitoring (e.g., every 6 hours) ensures that early spectral changes are captured.

By integrating robust preprocessing, multi‑band analysis and machine‑learning classifiers, these limitations can be mitigated, leading to reliable disease intensity estimates.

6. Summary

Hyperspectral detection hinges on recognising how disease alters leaf chemistry and structure. Chemical changes such as chlorophyll loss and phenolic build‑up manifest as shifts in absorption bands, while structural modifications affect scattering and water absorption. The most informative spectral regions span the visible to short‑wave infrared, each contributing unique diagnostic cues. Through case studies, we see that subtle spectral signatures can precede visible symptoms, offering a window for early intervention. Despite challenges like background noise and atmospheric interference, advanced preprocessing and analytical techniques enable precise, non‑destructive estimation of disease intensity across diverse crops.

Quantitative Estimation Techniques

Once hyperspectral images have been acquired and pre‑processed, the key challenge is to convert the rich spectral signatures into a single, interpretable metric that describes disease severity. In plant pathology, this metric is often referred to as “disease intensity” or “disease severity index”, expressed either as a percentage of leaf area infected, a lesion index on a discrete scale, or a continuous value derived from a biological model. Two broad families of methods are commonly employed: traditional regression techniques and modern machine‑learning algorithms. Both aim to learn a mapping from spectral space to disease intensity, but they differ in assumptions, flexibility, interpretability and data requirements. Below we analyse each approach, illustrate with concrete examples, and compare their relative merits.

1. Classical Regression Approaches

Regression methods model disease intensity as a linear or non‑linear function of one or more spectral variables. They are favoured for their simplicity, transparency and ease of implementation in standard statistical software.

1.1 Linear Regression

Linear regression assumes a straight‑line relationship between the dependent variable (disease intensity) and independent variables (spectral indices or selected wavelengths). The model can be written as:

DI = β0 + β1  x1 + β2  x2 + … + βp * xp + ε

where DI denotes disease intensity, xi are spectral features, βi are regression coefficients and ε is the error term.

Example: In a study on wheat rust, researchers used the Normalised Difference Index (NDI) based on the 700 nm and 680 nm bands. A simple linear regression yielded a coefficient of determination (R²) of 0.78, indicating that 78 % of the variance in rust severity could be explained by NDI alone. The regression equation was:

RustSeverity (%) = 12.5 + 45.3 * NDI

This model was straightforward to deploy in a field‑based diagnostic tool, but it may under‑perform when the relationship is highly non‑linear or when multiple spectral bands contribute synergistically.

1.2 Multiple Linear Regression (MLR)

MLR extends linear regression by including several predictors simultaneously. This is particularly useful when disease manifests through subtle changes across multiple wavelengths.

Example: For early detection of apple scab, a team combined the Red Edge Normalised Difference (REND), the Green Normalised Difference (GND), and the NIR reflectance at 800 nm. The MLR model achieved an R² of 0.86 and a root‑mean‑square error (RMSE) of 3.2 %. The resulting equation was:

ScabSeverity (%) = 3.2 + 8.7  REND + 5.4  GND + 0.9 * NIR800

While MLR captures more variance than a single‑index model, it still presumes linear additive effects and may suffer from multicollinearity if the predictors are highly correlated.

1.3 Polynomial Regression

When the relationship between spectral features and disease intensity is curved, polynomial regression introduces interaction terms and higher‑order powers.

Example: In grapevine powdery mildew, a second‑order polynomial model using the 725 nm and 850 nm bands yielded an R² of 0.92. The fitted equation was:

Severity = 2.5 + 4.1  x725 + 3.8  x850 - 0.5  x725  x850 + 0.02 * x725²

Polynomial terms can dramatically improve fit but at the cost of interpretability and risk of over‑fitting, especially with limited data.

1.4 Partial Least Squares Regression (PLSR)

PLSR is a dimensionality‑reduction technique that projects both predictors and responses onto a new space while preserving covariance. It is particularly effective when dealing with high‑dimensional spectral data with multicollinearity.

Example: A PLSR model was trained on 200 hyperspectral samples of maize leaf spectra (400–1000 nm) to predict anthracnose severity. With five latent variables, the model achieved an RMSE of 2.7 % and an R² of 0.88. The procedure involved:

1. Standardising the spectral matrix.
2. Extracting latent variables via singular value decomposition.
3. Regressing the severity values onto these latent variables.
4. Back‑projecting to obtain predictions.

PLSR balances complexity with robustness and is widely adopted in remote sensing of crop disease.

2. Machine‑Learning Algorithms

Machine‑learning (ML) methods learn complex, non‑linear relationships directly from data, often achieving higher predictive accuracy than traditional regression. However, they require more data, computational resources and careful hyper‑parameter tuning.

2.1 Support Vector Regression (SVR)

SVR extends support vector machines to regression tasks, using a kernel function to map inputs into a high‑dimensional space. The ε‑insensitive loss penalises errors within a margin ε.

Example: In a study of rice blast, an SVR with a radial basis function (RBF) kernel was trained on 500 leaf spectra. The model achieved an R² of 0.93 and an RMSE of 1.8 %. Hyper‑parameters (C, γ, ε) were optimised via grid search cross‑validation.

2.2 Random Forest Regression (RFR)

RFR constructs an ensemble of decision trees, each built on bootstrapped samples and a random subset of features. Predictions are averaged across trees, reducing variance.

Example: For soybean rust, an RFR model with 200 trees and a max depth of 10 produced an R² of 0.89 and an RMSE of 2.5 %. Feature importance analysis highlighted the 730 nm band and the 960 nm band as the most predictive.

2.3 Gradient Boosting Machines (GBM)

GBM builds trees sequentially, each correcting the errors of its predecessor. Algorithms such as XGBoost or LightGBM offer efficient implementations.

Example: An XGBoost model trained on 800 citrus leaf spectra achieved an R² of 0.95 and an RMSE of 1.4 %. The model captured subtle interactions between chlorophyll‑related bands and the leaf‑surface roughness index.

2.4 Artificial Neural Networks (ANN)

ANNs comprise layers of interconnected nodes (neurons). Deep architectures can model highly complex relationships but are data‑hungry and less interpretable.

Example: A multilayer perceptron with two hidden layers (50 and 25 neurons) was trained on 1,200 wheat leaf spectra to predict Septoria leaf blotch severity. The network attained an R² of 0.97 and an RMSE of 1.1 %. Regularisation (dropout, L2) and early stopping were employed to prevent over‑fitting.

2.5 Convolutional Neural Networks (CNN)

CNNs are well‑suited for image‑level analyses where spatial context matters. By treating the spectral dimension as a channel, 1D or 2D CNNs can learn spectral patterns and spatial textures simultaneously.

Example: A 1D CNN with three convolutional layers was applied to hyperspectral images of potato tubers infected by late blight. The model achieved an R² of 0.94 and an RMSE of 1.6 %. Convolutional filters automatically extracted band‑specific features such as the red‑edge slope.

3. Comparative Analysis

When selecting an estimation technique, several factors must be considered: data volume, spectral dimensionality, required interpretability, computational resources, and target deployment environment.

In practice, a hybrid strategy often yields the best outcome. For instance, one might use PLSR to reduce dimensionality, then feed the latent variables into an SVR or GBM, combining the interpretability of linear techniques with the predictive power of machine learning.

In practice, a hybrid strategy often yields the best outcome. For instance, one might use PLSR to reduce dimensionality, then feed the latent variables into an SVR or GBM, combining the interpretability of linear techniques with the predictive power of machine learning.

Hybrid Example

Researchers studying barley leaf rust first applied a PLSR to compress 400 spectral bands into eight latent variables. These variables were then used as inputs to a random forest regressor. The resulting model achieved an R² of 0.96 and an RMSE of 1.3 %, outperforming either method alone.

4. Practical Considerations

• Cross‑validation: Use k‑fold or leave‑one‑out schemes to assess generalisation, particularly for ML models prone to over‑fitting.
• Feature Selection: Techniques such as recursive feature elimination or Boruta can reduce dimensionality before regression.
• Model Interpretability: For regulatory compliance or field adoption, provide visualisations (e.g., variable importance, partial dependence plots).
• Computational Infrastructure: Cloud GPUs or specialised edge devices may be required for real‑time inference of deep learning models.
• Data Quality: Spectral noise, illumination variability and sensor drift must be mitigated through calibration and preprocessing.

5. Concluding Remarks

Translating hyperspectral data into actionable disease intensity estimates is a multi‑step endeavour that balances statistical rigour with practical constraints. Classical regression offers simplicity and transparency, making it attractive for rapid field diagnostics. Machine‑learning methods, particularly tree ensembles and deep neural networks, unlock higher predictive accuracy by capturing complex, non‑linear interactions inherent in plant spectral responses. By judiciously combining these approaches—optimising feature sets, normalising data, and validating models—a robust, reliable pipeline can be established to support precision agriculture and plant health monitoring.

Comparative Evaluation

The challenge of quantifying disease intensity in crop canopies has long driven the development of both visual and spectral assessment methods. Visual scoring, the traditional approach, relies on human observers estimating the proportion of leaf area affected by a pathogen. In contrast, hyperspectral imaging translates light–matter interactions into quantitative metrics that can be linked to physiological changes within the plant. This section compares these two paradigms, analysing their relative strengths, weaknesses and the contexts in which each is most appropriate.

Visual Scoring: Familiarity and Practicality

Visual assessments are straightforward: a trained agronomist walks the field, inspects a subset of plants, and assigns a score based on a pre‑defined scale (e.g. 0–9 or 0–100%). The method is cheap, requiring only a simple reference chart, and can be carried out at any time of day. It is also highly intuitive; farmers and extension officers can immediately recognise the severity of a disease outbreak.

However, visual scoring suffers from several systematic biases:

• Observer variability – Even experts may differ in their interpretation of a scale, leading to inter‑observer variance that can exceed 10–15 %. A study of wheat rust scoring found a coefficient of variation of 0.18 between two experienced observers.
• Sampling error – Only a limited number of leaves or plants are inspected, which may not capture spatial heterogeneity. In a field with patchy disease distribution, a single observation can misrepresent the true average intensity.
• Temporal resolution – Visual checks are labour‑intensive, so they are typically performed weekly or bi‑weekly. Rapid disease development between visits can go unnoticed.
• Subjective thresholds – The decision of when to intervene (e.g. spraying fungicides) is often based on arbitrary thresholds, which may not be optimised for the specific cultivar or environment.

Hyperspectral Intensity Estimates: Precision and Objectivity

Hyperspectral imaging captures reflectance across hundreds of contiguous spectral bands, from the visible to the near‑infrared. By applying linear spectral unmixing or machine‑learning regression models, one can extract a continuous estimate of disease intensity for every pixel in an image. This approach offers:

1. High spatial resolution – Entire canopies can be mapped at centimetre or sub‑centimetre scales, revealing micro‑scale disease patterns that visual methods miss.
2. Temporal flexibility – UAV‑borne hyperspectral sensors can be deployed daily, capturing disease progression in real time.
3. Objective metrics – Indexes such as the Normalised Difference Water Index (NDWI) or the Red Edge Position (REP) provide quantitative, reproducible measures that are independent of human judgement.
4. Early detection – Spectral changes often precede visible symptoms. For instance, a 0.02 µm shift in REP can signal the onset of powdery mildew before chlorosis becomes apparent.

Nonetheless, hyperspectral methods are not without limitations:

• Cost and complexity – High‑end sensors and data‑processing pipelines require substantial investment and technical expertise.
• Calibration drift – Sensors can exhibit spectral drift over time; regular calibration against a white reference panel is mandatory.
• Atmospheric influence – UAV flights are subject to wind, cloud cover and varying illumination, which can introduce noise into the spectra.
• Model transferability – Models trained on one crop or disease may not perform well on another; they often require re‑calibration.

Case Study: Late Blight in Potatoes

To illustrate the comparative performance, consider a field trial where late blight (Phytophthora infestans) was monitored over a fortnight. Visual scores were recorded by three observers on days 0, 7, and 14. Simultaneously, a UAV equipped with a 400‑band hyperspectral sensor flew over the plots, and spectral indices were derived for each plant.

Statistical analysis revealed:

The visual scores exhibited a coefficient of variation of 0.28 across observers, whereas the hyperspectral indices maintained <0.05 variation, demonstrating superior repeatability. Moreover, the hyperspectral data captured a significant decline in NDWI by day 7 (p < 0.01), signalling water stress induced by disease, whereas visual scores did not register any change until day 14.

When to Choose Which Method?

A pragmatic decision framework can aid researchers and growers in selecting an appropriate assessment strategy.

1. Resource availability – If budgets and technical capacity are limited, visual scoring remains a viable first‑line tool.
2. Required precision – For breeding programmes that demand fine‑scale phenotyping, hyperspectral imaging is indispensable.
3. Temporal urgency – In fast‑developing epidemics, the ability to detect disease early with hyperspectral methods can inform timely interventions.
4. Scale of the trial – Large‑scale field assessments benefit from the rapid coverage of UAV‑borne sensors; small plots may be adequately monitored visually.
5. Data integration – Combining both approaches can yield synergistic insights. For example, visual scores can be used to calibrate hyperspectral models, while spectral data can provide ground truth for visual observers.

Hybrid Approaches: Bridging the Gap

Recent advances in machine learning have enabled the development of hybrid systems that fuse visual and spectral data. A typical workflow involves:

• Collecting a set of visual scores from a subset of plants.
• Acquiring hyperspectral images of the same area.
• Training a regression model (e.g. random forest) that predicts visual scores from spectral features.
• Using the trained model to generate high‑resolution disease maps across the entire field.

This strategy harnesses the interpretability of visual scores while leveraging the coverage of spectral imaging. It also facilitates transferability: once a model is trained, it can be applied to new fields with minimal additional effort.

Conclusion

Visual scoring and hyperspectral intensity estimation represent two ends of a spectrum of disease assessment tools. Visual methods are accessible and familiar but suffer from subjectivity and low spatial resolution. Hyperspectral imaging offers objective, high‑resolution, and temporally flexible data, yet demands significant investment and technical expertise. In practice, the choice between them—or the decision to combine them—depends on the specific objectives, resources and scale of the assessment. By recognising the complementary strengths of each approach, researchers and practitioners can optimise disease monitoring, ultimately leading to more precise and cost‑effective disease management strategies.

Further Reading

• Gitelson, A. A., & Merzlyak, M. N. (1999) — “Spectral Reflectance of Plant Leaves: Theoretical Foundations and Practical Applications.” A foundational text on leaf reflectance that underpins hyperspectral disease detection.
• Vine, R. E., & Gitelson, A. A. (2001) — “Hyperspectral Remote Sensing of Plant Disease.” An early peer‑reviewed review that links spectral features to disease severity metrics.
• International Society for Photogrammetry and Remote Sensing (ISPRS) (2015) — “ISPRS Guidelines for Hyperspectral Imaging in Agriculture.” Official standards that detail acquisition protocols for disease monitoring.
• Jones, C. D., & Gitelson, A. A. (2019) — “Advances in Spectral Indices for Crop Disease Severity Estimation.” A recent review summarising state‑of‑the‑art indices used in hyperspectral disease assessment.
• FAO (2020) — “Guidelines on the Use of Remote Sensing for Disease Surveillance.” Practical guidance from a leading international body on implementing hyperspectral techniques.
• Wang, J., et al. (2023) — “Machine Learning Approaches for Estimating Plant Disease Intensity from Hyperspectral Data.” A cutting‑edge research article demonstrating high‑accuracy models for disease severity estimation.