Lesson 8.1.1: Classification versus severity prediction

1. Foundations of Classification and Severity Prediction

In disease severity estimation, two analytical approaches dominate: classification and severity prediction. Although they share the same overarching goal – to inform clinical decision‑making – their methodological foundations, output formats and typical use‑cases differ substantially. Understanding these differences is essential for selecting the appropriate model for a given health‑care context.

1.1 What is Classification?

Classification is a form of supervised learning where the response variable is categorical. In medical contexts this often means assigning a patient to one of several discrete risk groups, such as:

• Low, moderate or high risk of developing a complication.
• Presence or absence of a particular comorbidity.
• Responder vs. non‑responder to a treatment regimen.

Typical algorithms include logistic regression, decision trees (CART), random forests, support vector machines and neural networks. The key property of a classification model is that it produces a probability or a hard class label. For example, a logistic regression model might output that a patient has a 0.73 probability of progressing to severe disease, and a clinician can decide to treat the patient as high risk if the probability exceeds a chosen threshold.

1.2 What is Severity Prediction?

Severity prediction, often referred to as regression or ordinal regression in medical research, treats the outcome as a continuous or ordered variable. Instead of merely saying “high risk” or “low risk,” the model estimates a numerical severity score or a continuous risk estimate. Example outputs include:

• Estimated oxygen saturation deficit measured in percentage points.
• Predicted length of stay in days.
• Estimated disease‑specific severity index such as the APACHE II score.

Common algorithms are linear regression, ridge or lasso regression, Gaussian process regression, and more recently, deep learning models that output a scalar value. Because the output is continuous, severity prediction allows finer-grained resource allocation – for instance, deciding whether to admit a patient to a high‑dependency unit or to a general ward.

1.3 Key Differences in Purpose

While both approaches aim to inform clinical pathways, their purposes diverge:

1. Decision Thresholds – Classification models naturally lend themselves to binary decisions (e.g. “treat” vs. “observe”). Severity prediction requires setting a continuous threshold, which may be less intuitive for clinicians.
2. Granularity of Information – Severity prediction delivers a spectrum of risk, allowing nuanced triage. Classification collapses this spectrum into broad categories, potentially oversimplifying complex clinical realities.
3. Interpretability – Simple classification rules (e.g. decision trees) can be easily explained to clinicians. Severity prediction models, particularly complex regressions or neural networks, may be less transparent, though the numeric output can be more directly interpretable clinically.
4. Data Requirements – Classification can often perform well with limited data, especially if the outcome classes are imbalanced but well‑represented. Severity prediction typically requires larger datasets to estimate the continuous outcome accurately.

1.4 Typical Use‑Cases

Below are concrete examples illustrating where each approach is most appropriate.

1.4.1 Classification Use‑Cases

• Sepsis Early Warning – Hospitals use sepsis screening tools that classify patients as “sepsis” or “no sepsis” based on vital signs. A binary decision is required for rapid intervention.
• COVID‑19 triage – Classifying patients into “hospitalised” or “non‑hospitalised” risk groups to determine admission thresholds.
• Oncology – Determining whether a patient is a candidate for a specific targeted therapy based on biomarker presence (positive/negative).

1.4.2 Severity Prediction Use‑Cases

• Acute respiratory distress syndrome (ARDS) – Predicting the PaO₂/FiO₂ ratio to guide ventilatory support intensity.
• Chronic kidney disease – Estimating glomerular filtration rate (GFR) as a continuous outcome to stage disease severity.
• Cardiac event risk – Calculating a continuous risk score (e.g. 10‑year cardiovascular risk) rather than a binary “high risk” label.

1.5 Comparing Alternatives

When deciding between classification and severity prediction, consider the following dimensions:

1.6 Practical Considerations

1. Evaluation Metrics – Classification is evaluated with accuracy, sensitivity, specificity, AUC‑ROC, or F1‑score. Severity prediction uses mean squared error (MSE), mean absolute error (MAE), or R². Choosing the metric that aligns with clinical priorities is crucial.

2. Imbalanced Data – Classification often faces class imbalance. Techniques such as SMOTE, class weighting or focal loss can mitigate bias. Severity prediction is less sensitive to imbalance but can suffer from heteroscedasticity, requiring robust regression methods.

3. Calibration – For classification, probability calibration (Platt scaling, isotonic regression) ensures that predicted probabilities reflect true risk. In severity prediction, residual analysis and heteroscedasticity checks confirm model reliability.

4. Regulatory and Ethical Issues – Classification models can be perceived as black boxes if complex. Severity prediction models that output a single risk score may be more acceptable in regulatory filings, provided the score is well‑validated.

1.7 Conclusion

Classification and severity prediction are complementary tools in the armamentarium of disease severity estimation. Classification excels when rapid, binary decisions are required, whereas severity prediction offers a nuanced, continuous risk assessment that supports sophisticated resource planning. Selecting the appropriate approach hinges on the specific clinical question, data availability and the desired granularity of the outcome. A thoughtful alignment of methodological choice with clinical workflow ensures that predictive analytics translate into meaningful patient outcomes.

2. Mechanistic Distinctions: Algorithms and Data Requirements

In the context of medical risk assessment, the distinction between a classification model—predicting a binary outcome such as “disease present” versus “absent”—and a severity prediction model—estimating a continuous or ordinal risk score—has profound implications for the choice of algorithm, the handling of features, and the structure of the data that feeds into the model. Below we analyse these differences in detail, drawing on concrete examples from cardiovascular and oncological risk prediction.

2.1 Algorithmic Families: Discriminative vs. Predictive Modelling

Classification tasks are typically addressed with discriminative algorithms that maximise the separation between classes. Logistic regression, support vector machines (SVMs), decision trees and their ensembles (random forests, gradient boosting machines) are the most common choices. These methods implicitly assume that the data are drawn from two distinct distributions and focus on learning a decision boundary.

Severity prediction, on the other hand, is a regression problem. Algorithms such as linear regression, Poisson regression, Cox proportional hazards models, or Bayesian hierarchical models are often employed. When the outcome is ordinal (e.g. stages I–IV cancer), ordinal logistic regression or proportional odds models are suitable. In both cases, the objective is to approximate a mapping from feature space to a continuous score rather than a binary label.

Because the loss functions differ—cross‑entropy or hinge loss for classification versus mean squared error or negative log‑likelihood for regression—regularisation, optimisation dynamics and convergence characteristics vary. For instance, logistic regression benefits from L2 regularisation to prevent overfitting when predictors are highly collinear, whereas ridge regression (an L2‑regularised linear model) is the natural analogue for severity prediction.

2.2 Feature Handling: Discretisation, Interaction and Hierarchy

Both classification and severity models can incorporate categorical variables. However, the treatment of interactions and higher‑order terms often differs. In a classification setting, a decision tree may automatically capture interactions by splitting on different features along a path, whereas in a regression context we usually need to specify interaction terms explicitly or rely on polynomial kernels in SVMs.

When dealing with ordinal outcomes, preserving the inherent ordering is crucial. Ordinal logistic regression enforces a monotonic relationship between predictors and the cumulative logits. In contrast, a standard binary classifier would ignore this structure, potentially yielding inconsistent probability estimates. This is illustrated by the following snippet:

# Ordinal logistic regression using scikit‑learn
from mord import LogisticIT
model = LogisticIT()
model.fit(X, y_ordinal)

Here, y_ordinal might represent cancer stages 0–4. The model learns a single set of coefficients shared across thresholds, ensuring that higher stages receive higher predicted probabilities.

Feature scaling also plays a different role. For SVMs in classification, kernel choice and scaling can dramatically affect decision boundaries. In severity prediction using linear regression, scaling primarily influences numerical stability but does not alter the underlying relationship.

2.3 Data Structure and Sample Size Requirements

Classification models can often tolerate smaller datasets if the classes are well separated. However, severe class imbalance—common in rare disease screening—can lead to biased predictions. Techniques such as SMOTE (Synthetic Minority Over-sampling Technique) or class‑weight adjustments are routinely applied. Severity prediction usually relies on larger sample sizes to estimate continuous risk accurately, especially when the outcome variance is high. For example, predicting 10‑year mortality risk with a Cox model typically requires thousands of events to achieve stable hazard ratios.

In addition, severity models often involve censoring, as in survival analysis. This introduces a data structure where the outcome is partially observed, necessitating specialised likelihood functions. Classification models rarely encounter censoring; missing labels are usually treated as absent cases or imputed.

Consider the following table summarising typical data requirements:

2.4 Model Evaluation Metrics and Their Implications

Evaluation metrics differ markedly. Classification models are judged by accuracy, sensitivity, specificity, area under the receiver operating characteristic curve (AUC‑ROC) and calibration plots. Severity models use metrics such as mean absolute error (MAE), root mean squared error (RMSE), R², concordance index (C‑index) for survival, and calibration of predicted probabilities against observed outcomes.

These metric choices influence algorithm selection. For instance, a random forest tuned to maximise AUC may produce highly discriminatory but poorly calibrated probabilities, which is undesirable in a clinical risk score. Conversely, a penalised Cox model may achieve a lower C‑index but provide more reliable hazard ratios for policy decisions.

2.5 Practical Example: Cardiovascular Risk Prediction

Take the widely used Framingham risk score. The original model is a linear regression predicting 10‑year absolute risk of coronary heart disease. It incorporates age, cholesterol, blood pressure, smoking status and diabetes as predictors. To derive a probability, the linear predictor is passed through a logistic function:

lin_pred <- β0 + β1age + β2chol + β3bp + β4smoke + β5*diabetes
prob <- 1 / (1 + exp(-lin_pred))

Here, the outcome is continuous (probability) rather than binary. If we re‑formulate the problem as a classification task—predicting whether the 10‑year risk exceeds a threshold (e.g. 20%)—we would train a logistic regression directly on a binary label. The resulting model would optimise classification accuracy but might lose nuance in risk stratification.

In contrast, a severity model for predicting tumour size would use a linear or non‑linear regression, perhaps incorporating imaging features extracted via convolutional neural networks (CNNs). The CNN would output a feature vector that the regression layer maps to a size estimate. The loss function would be mean squared error, not cross‑entropy, and the model would be tuned to minimise prediction error rather than misclassification.

2.6 The Role of Domain Knowledge and Feature Engineering

Both types of models benefit from expert knowledge. However, severity prediction often requires more sophisticated feature engineering to capture subtle variations. For example, in predicting disease progression, longitudinal biomarkers and time‑varying covariates are incorporated via time‑dependent covariate models or deep learning architectures like LSTM networks. Classification models may simply summarise these signals (e.g. mean, variance) before feeding them to a tree‑based algorithm.

Moreover, severity models frequently rely on hierarchical data structures—patients nested within clinics, repeated measurements over time. Mixed‑effects models or hierarchical Bayesian approaches can explicitly model these dependencies, improving predictive accuracy and generalisability.

2.7 Regularisation and Model Complexity

Classification tasks with high-dimensional feature spaces (e.g. genomics) often require strong regularisation to avoid overfitting. Techniques such as LASSO (L1) or elastic net combine variable selection with shrinkage, producing sparse models that are easier to interpret clinically. Severity prediction also benefits from regularisation, but the penalty may be tuned differently. For instance, ridge regression (L2) is preferred when multicollinearity is present but variable selection is not the primary goal.

When deploying models in a clinical setting, the trade‑off between bias and variance must be carefully managed. Overly complex severity models may achieve low training error but perform poorly on external validation. Conversely, a simplistic classification model may be robust but miss subtle risk gradients.

2.8 Summary

The choice between classification and severity prediction dictates the entire modelling pipeline: from algorithm selection and loss function to data preprocessing, feature representation and evaluation metrics. Classification models optimise discriminatory power between discrete classes, often using tree‑based or support‑vector approaches, and handle missing data primarily through imputation or class‑weighting. Severity models aim to approximate a continuous risk surface, employing regression or survival techniques that account for censoring and longitudinal structure. Understanding these mechanistic distinctions is essential for building clinically useful risk tools that are both accurate and interpretable.

3. Clinical and Operational Implications

In clinical practice, the distinction between classification and severity prediction is not merely an academic exercise; it shapes every downstream decision, from the choice of therapeutic intervention to the allocation of scarce resources such as intensive care beds. Below we examine how each modelling paradigm influences treatment pathways, operational logistics, and ultimately patient outcomes, drawing on concrete case studies to illustrate the practical consequences.

3.1 Classification Models: Binary Decision-Making

Classification algorithms produce a discrete label – for instance, “disease present” versus “disease absent.” In the context of a chronic illness such as type 2 diabetes, a classification model might flag patients as either high‑risk for complications or low‑risk. This binary output allows clinicians to trigger predefined management protocols:

• Patients classified as high‑risk receive an intensified monitoring schedule, early referral to a specialist, and possibly an escalation of pharmacotherapy.
• Low‑risk patients remain on standard care pathways, reducing unnecessary clinic visits and medication costs.

Because the decision boundary is sharp, classification models are well‑suited for triage situations where a rapid, clear action is required. For example, in emergency departments, a classification model that predicts sepsis can prompt immediate antibiotic administration, potentially saving lives. The operational benefit lies in the reduction of decision time and the standardisation of care pathways.

3.2 Severity Prediction Models: Quantitative Risk Assessment

Severity prediction models, often expressed as probability scores or continuous risk indices, provide a spectrum of risk rather than a single label. In the same diabetes example, a severity model might estimate the probability that a patient will develop nephropathy within the next five years (e.g., 12 % versus 45 %). This nuanced information supports shared decision‑making and personalised care plans.

Operationally, severity scores enable resource optimisation. Hospitals can prioritise ambulatory care teams or allocate specialist consultations to those with the highest predicted risk, thereby improving the utilisation of limited specialist time. In oncology, a severity score predicting tumour aggressiveness can inform the choice between active surveillance and immediate intervention.

3.3 Impact on Treatment Decisions

When a model outputs a binary classification, treatment decisions follow a “one‑size‑fits‑all” logic. However, the lack of granularity can lead to over‑ or under‑treatment. Consider the case of atrial fibrillation (AF) risk scoring: a classification model that simply flags patients as “high risk for stroke” may trigger anticoagulation in all flagged patients, including those with a low absolute risk of bleeding. A severity model that estimates a 2 % versus 10 % annual stroke risk allows clinicians to weigh the benefits of anticoagulation against bleeding risk, tailoring therapy accordingly.

Severity predictions also influence drug dosing. In pharmacogenomics, a severity model that predicts the likelihood of a severe adverse reaction to a particular medication informs whether a patient should receive a standard dose, a reduced dose, or an alternative drug altogether. This precision reduces adverse events and improves adherence.

3.4 Resource Allocation in Healthcare Settings

From an operational perspective, the choice between classification and severity prediction has profound implications for resource planning:

1. Bed Management – In intensive care units (ICU), a classification model that predicts ventilator requirement can trigger early mobilisation of ICU beds. Conversely, a severity model that estimates the probability of prolonged ventilation allows administrators to anticipate staffing needs and equipment allocation more accurately.
2. Staffing and Training – For chronic disease management programmes, a severity score can inform the allocation of nurse‑led follow‑up visits. High‑severity patients may receive more frequent contact, while low‑severity patients are managed through telehealth platforms, optimising workforce utilisation.
3. Supply Chain Management – Pharmacy departments can use severity predictions to forecast medication stock levels. For instance, a higher predicted risk of osteoporosis in post‑menopausal women can trigger increased procurement of bisphosphonates.

3.5 Patient Outcomes and Satisfaction

Patient outcomes are directly affected by the granularity of risk information. A classification approach may lead to overtreatment, which can cause unnecessary side effects, anxiety, and financial burden. For example, a blanket recommendation for prophylactic anticoagulation in all patients flagged as high risk for atrial fibrillation may expose low‑risk patients to an elevated bleeding risk without commensurate benefit.

Severity models, by contrast, support shared decision‑making and foster patient engagement. When patients understand that their individual risk of disease progression is modest, they may be more willing to adhere to lifestyle modifications rather than immediately accept invasive interventions. This patient‑centred approach has been shown to improve long‑term outcomes in conditions such as hypertension and hyperlipidaemia.

Moreover, severity predictions can reduce health disparities. By providing objective risk estimates, clinicians can avoid biases that may arise when relying solely on clinical judgement, which is often influenced by socioeconomic factors.

3.6 Model Calibration and Clinical Trust

Regardless of the paradigm, model calibration is essential for clinical trust. A poorly calibrated classification model may misclassify a significant proportion of patients, leading to inappropriate treatment. Severity models, if over‑ or under‑estimating risk probabilities, can misguide resource allocation. For instance, a severity model that overestimates the risk of malignancy in lung nodules may prompt unnecessary biopsies, increasing patient anxiety and healthcare costs.

Regular audit and recalibration using up‑to‑date data are therefore mandatory. In practice, hybrid approaches that combine classification thresholds with severity scores often yield the best balance between simplicity and precision.

3.7 Case Study: COVID‑19 Hospitalisation Risk

During the early phases of the COVID‑19 pandemic, several institutions implemented classification models that flagged patients as “likely to hospitalise” versus “unlikely to hospitalise.” This binary output enabled rapid triage but resulted in a high rate of false positives, overwhelming hospital capacity.

Subsequent models shifted to severity prediction, estimating the probability of requiring intensive care. Hospitals could then allocate ICU beds to those with the highest predicted needs, improving utilisation and reducing mortality. This example underscores how the granularity of risk information directly influences operational efficiency and patient outcomes.

3.8 Recommendations for Clinical Implementation

• Choose classification models for high‑stakes, low‑time‑budget scenarios (e.g., emergency triage).
• Adopt severity prediction models when personalised treatment plans and resource optimisation are paramount.
• Implement hybrid systems that trigger a severity score only when classification flags a patient as high risk.
• Ensure continuous model monitoring, calibration, and clinician engagement to maintain trust and effectiveness.

In summary, classification and severity prediction models serve distinct yet complementary roles in modern healthcare. Their appropriate application can streamline clinical workflows, optimise resource utilisation, and, most importantly, improve patient outcomes by aligning treatment intensity with individual risk profiles.

4. Comparative Evaluation and Decision Guidance

When developing a decision‑support system for healthcare, one of the first choices clinicians and data scientists must grapple with is whether to frame the problem as a classification task or a severity prediction task. The distinction is subtle yet profound: classification seeks a discrete outcome (e.g. “disease present” vs “disease absent”), whereas severity prediction estimates a continuous or ordinal measure of disease burden (e.g. pain score, organ dysfunction score). The following framework illustrates how to weigh the strengths, limitations and suitability of each approach in different clinical contexts.

4.1 Decision Matrix: Classification vs Severity Prediction

Below is a concise decision matrix that maps key clinical attributes against the two modelling paradigms. Use it as a quick reference when confronted with a new problem statement.

4.2 Strengths and Limitations

4.2.1 Classification

Strengths

• Intuitive interpretation: “positive” vs “negative” outcomes align with clinical decision thresholds.
• Typically requires less data to train; robust to noise in labels.
• Decision rules can be expressed as simple if‑then statements, facilitating transparency.
• Well‑established performance metrics (accuracy, AUC‑ROC, F1‑score) aid comparison across studies.

Limitations

• Loss of information when discretising continuous variables (e.g. a patient with a marginally abnormal lab is treated the same as one with a markedly abnormal value).
• Cannot provide a calibrated probability of severity; may lead to over‑ or under‑treatment.
• Binary outputs may not respect clinical nuance, especially in conditions with a spectrum of presentations.

4.2.2 Severity Prediction

Strengths

• Preserves the full spectrum of disease burden; supports personalised medicine.
• Enables risk‑adjusted resource planning (e.g. predicting ventilator days).
• Facilitates dose‑response studies and threshold optimisation.
• Often yields better calibration, allowing clinicians to quantify uncertainty.

Limitations

• Requires high‑quality, granular outcome data; susceptible to measurement error.
• Model outputs can be harder to explain, potentially reducing clinician trust.
• Continuous predictions may be misused if thresholds are applied incorrectly.
• Regulatory approval may be more demanding due to the need for rigorous validation.

4.3 Concrete Clinical Scenarios

4.3.1 Acute Myocardial Infarction (AMI)

In the emergency department, a rapid decision is needed: should a patient receive thrombolysis? The classification approach—predicting “thrombolysis‑benefit” vs “no benefit”—is favoured because it aligns with the time‑critical nature of the intervention. Studies have shown that a binary risk score (e.g. TIMI risk score) can be calculated in under a minute and directly informs treatment protocols.

Conversely, if the aim is to optimise hospital bed utilisation post‑AMI, a severity prediction model estimating expected length of stay and need for cardiac rehabilitation can be invaluable. Such a model can be built on continuous variables (e.g. troponin levels, ejection fraction) and output a predicted LOS distribution, allowing planners to allocate resources more efficiently.

4.3.2 Chronic Obstructive Pulmonary Disease (COPD) Exacerbations

For patients presenting with an exacerbation, clinicians often use the Modified Medical Research Council Dyspnoea Scale (mMRC) to gauge severity. A severity prediction model that estimates the probability of requiring ICU admission based on continuous spirometry values and arterial blood gas results provides a nuanced risk profile. This is preferable to a simple “needs ICU” vs “does not need ICU” classification because it captures the gradation of respiratory failure.

However, when triaging patients in a busy emergency department, a quick classification tool that flags those at high risk of imminent respiratory arrest (e.g. using the CURB‑65 score adapted for COPD) can expedite life‑saving interventions. The choice hinges on the immediacy of the decision and the availability of detailed physiological data.

4.3.3 Breast Cancer Screening

In mammography, a binary “malignant” vs “benign” classification is the standard for radiologists. Nonetheless, severity prediction models that estimate tumour aggressiveness (e.g. Ki‑67 proliferation index) can inform personalised treatment plans. A hybrid approach is often used: a classification model first flags suspicious lesions, and a severity model then stratifies lesions into low‑, intermediate‑, and high‑risk categories for biopsy prioritisation.

4.4 Implementation Considerations

4.4.1 Data Quality and Labelling

Classification models can tolerate noisy labels because the decision boundary is coarse. Severity prediction, however, is highly sensitive to label precision. In practice, this means:

• Ensuring continuous outcomes are measured with standardised instruments.
• Applying rigorous quality control to avoid systematic bias.
• Considering semi‑supervised or transfer learning techniques when continuous labels are scarce.

4.4.2 Model Explainability

Regulatory bodies increasingly demand that predictive models be interpretable. For classification, simple decision trees or rule‑based systems often suffice. For severity prediction, techniques such as SHAP values or partial dependence plots can elucidate feature contributions, but the complexity can still hinder clinician acceptance.

4.4.3 Calibration and Threshold Selection

Even a perfect classifier can be miscalibrated if its probability estimates are systematically high or low. Severity prediction models usually produce probability distributions that can be calibrated with Platt scaling or isotonic regression. Selecting appropriate thresholds (for example, deciding that a predicted severity score above 0.8 warrants aggressive treatment) requires collaboration with clinicians and may involve cost‑benefit analyses.

4.5 Decision Guidance Checklist

1. Define the clinical question. Is the priority a yes/no decision or a graded risk estimate?
2. Assess data availability. Do you have reliable continuous outcomes? If not, classification may be safer.
3. Consider workflow integration. Can the model output be delivered in the required time frame?
4. Evaluate regulatory implications. Are there existing guidelines for the chosen paradigm?
5. Plan for explainability. Will the output be understandable to end‑users?
6. Validate rigorously. Use cross‑validation, external datasets and, where possible, prospective studies.

By systematically applying this framework, clinicians and data scientists can make informed choices about whether to pursue a classification or severity prediction model, thereby aligning technical solutions with clinical needs and organisational constraints.

Further Reading

• World Health Organisation (2023) — WHO Clinical Management of COVID-19. Provides a framework for classifying disease severity and guiding treatment decisions.
• American College of Chest Physicians (2021) — Chest Guidelines for the Management of Acute Respiratory Distress Syndrome. Details severity grading and evidence‑based therapeutic thresholds.
• Goodman, D. & Harris, A. (2022) — Clinical Epidemiology of Infectious Diseases. Chapter on severity prediction models illustrates statistical methods for risk stratification.
• European Centre for Disease Prevention and Control (2020) — ECDC Guidance on Severity Assessment of Emerging Infectious Diseases. Offers standardised criteria used across EU health systems.
• Harris, J., Ahmed, R. & Kumar, S. (2024) — Machine Learning for Disease Severity Estimation. Reviews current algorithms and validation practices in clinical datasets.