Literature Review

2026-03-17

The 12 Enemies of Clinical Prediction Models

This paper distills decades of experience in clinical epidemiology and prediction modeling into a set of recurring problems, provocatively called “enemies”, that undermine prediction models long before they ever reach the bedside.

Reading the paper resonated strongly with concerns I raised in an earlier blog post about AI/ML in medicine. In many projects, we spend enormous effort asking how to build models, while the important question of why this model should exist at all is often overlooked.

What this paper shows clearly is that the biggest obstacles to reliable clinical AI are rarely algorithmic.

Most problems appear before the model is even trained.

The Real Problem: Many Models, Little Clinical Impact

Prediction modeling is one of the most active areas of biomedical research. New models are proposed for disease diagnosis, prognosis prediction, high-risk patient identification, and treatment recommendation. However, systematic reviews consistently reveal a troubling pattern:

• many models are developed

• few are externally validated

• even fewer are evaluated for clinical impact

• almost none become routine tools in healthcare

In other words, the field produces many models but relatively little usable knowledge.

To explain why, the authors organize the challenges into twelve “enemies,” grouped into four broad categories:

• the nature of clinical problems

• the quality of available data 

• methodological choices in model development

• the culture and incentives of biomedical research

Together, these enemies illustrate why reliable prediction modeling is much harder than simply applying machine learning to medical data.

Enemy 1: Medical Data Often Contain Weak Signals

Biological systems are highly complex. Most diseases arise from multiple interacting pathways involving genetics, environment, lifestyle, and chance. As a result, predictors rarely cleanly separate patients into clear outcome groups. For example, two patients with identical risk factors for cardiovascular disease may still experience very different outcomes. In such settings, flexible machine learning models can easily mistake random noise for a meaningful signal, especially when datasets are small. Interestingly, systematic comparisons often show that complex machine learning models do not consistently outperform simpler statistical models in clinical prediction tasks. More sophisticated algorithms cannot compensate for weak underlying signals.

Enemy 2: Individual Risk Is Inherently Uncertain

Prediction models frequently generate individualized risk estimates. For example: “Your risk of developing disease X within five years is 20%.”

However, a patient belongs to many possible reference groups depending on which predictors are considered. Two equally valid models might assign very different risk scores to the same individual while still performing similarly overall.

This highlights an often overlooked limitation: prediction models estimate probabilities, not certainties.

Enemy 3: Healthcare Systems Are Not Stable

Clinical environments vary across hospitals, countries, and time. A model trained in one setting may perform poorly in another because:

• patient populations differ

• diagnostic practices vary

• treatments evolve

• measurement procedures change

For example, a model developed using laboratory data from one hospital may rely on tests that are measured differently elsewhere. Even within the same institution, models may degrade over time as clinical practices change, a phenomenon sometimes called data drift.

Enemy 4: Clinical Data Are Messy

 In medicine, the biggest challenge is often data quality. Electronic health records contain missing values, inconsistent coding, irregular measurements, and complex clinical workflows. For example, certain laboratory tests are only ordered when physicians suspect severe disease. When those tests appear “missing” in other patients, the missingness itself reflects clinical decision-making rather than random absence. Ignoring this structure can distort model predictions.

Enemy 5: Diagnostic Labels Are Imperfect

Prediction models rely on labeled outcomes. Yet in medicine, diagnostic labels are often uncertain. Some diseases can only be confirmed through selective, invasive procedures. Others depend on subjective clinical judgment or evolving diagnostic criteria. As a result, models may be trained on labels that are themselves imperfect approximations of reality.

Enemy 6: Treatment Can Change the Outcome

Prediction models can influence clinical decisions. This creates a paradox.

Suppose a model predicts that a patient has a high risk of a heart attack. Doctors intervene with preventive treatment, and the heart attack never occurs. When evaluated retrospectively, the model may appear inaccurate even though it helped prevent the event. This phenomenon is sometimes called the treatment paradox.

Prediction models interact with clinical decisions, which means they can change the very outcomes they aim to predict.

Enemy 7: Methodological Shortcuts

Many prediction models suffer from avoidable methodological problems. Common examples include:

• small datasets with few outcome events

• excessive model complexity

• inappropriate variable selection

• arbitrary categorization of continuous predictors. For instance, converting continuous variables such as age or blood pressure into simple thresholds (“above 60” vs “below 60”) discards valuable information.

Such practices can produce models that appear accurate during development but fail during validation.

Enemy 8: Poor Reporting

Even well-developed models become difficult to use when poorly reported. Researchers sometimes fail to publish:

• the full model equation

• predictor definitions

• implementation details

Without these elements, other researchers cannot validate or apply the model. Scientific progress depends on reproducibility.

Enemy 9: Clinical Impact Is Rarely Studied

A model might accurately predict outcomes yet fail to improve patient care if clinicians do not use it or if it does not alter decisions.

Evaluating clinical impact requires careful studies, sometimes randomized trials, to determine whether model use actually improves outcomes. Such studies are rare.

Enemy 10: Scientific Incentives Favor Novelty

Academic incentives reward new discoveries. Publishing a new model is often easier than validating an existing one. As a result, the literature accumulates hundreds of similar models for the same clinical problem while rigorous validation studies remain scarce. The authors argue that the field would benefit from fewer models and more validation.

Enemy 11: Data Science Education Can Be Too Narrow

Many data science training programs emphasize algorithms and programming, but devote less attention to:

• study design

• domain knowledge

• causal reasoning

• clinical context

Students learn how to build models, but not always when the act of building a model is scientifically justified. Prediction modeling is not purely a computational activity; it is part of clinical research.

Enemy 12: Commercial Opacity

Healthcare AI is increasingly commercialized. Some companies treat predictive models as proprietary products, restricting access to their algorithms or training data. Without transparency, independent validation becomes difficult. In several cases where independent evaluations were conducted, commercial models performed significantly worse than advertised.

Why These Lessons Matter Now

These issues are not merely theoretical. Hospitals around the world are actively exploring how to integrate AI into clinical workflows, from triage systems to diagnostic support tools. However, deploying models without addressing the challenges described above risks introducing unreliable tools into healthcare.

The key lesson from the prediction modeling literature is simple:

The biggest risks in clinical AI rarely come from algorithms.
They come from poorly defined problems, weak datasets, and inadequate validation.

Beyond Hospitals: Consumer Health and Omics Platforms

These challenges also apply to companies offering predictive health insights based on biological measurements.

Many platforms now analyze:

• genomic variants

• metabolomic profiles

• microbiome composition

• multi-omics data

These services often promise to estimate health risks or provide personalized recommendations. Such predictions rely on models trained on datasets linking biological measurements to health outcomes. For genomic platforms, predictors are largely static. Even there, however, models must contend with population differences and uncertain interpretation of genetic variants.

For metabolomics or microbiome-based services, the challenges may be even greater. These biological signals vary over time and are influenced by diet, medications, environment, and lifestyle.

As a result, many of the “enemies” described earlier appear in this space as well:

• weak signal-to-noise ratios

• heterogeneous populations

• uncertain labels

• limited external validation

This does not mean these technologies lack promise. But it does mean that strong predictive claims require strong evidence.

What Good Prediction Model Research Should Look Like

If so many models fail, what distinguishes good prediction modeling? Several principles emerge from the literature.

1. First, models should start with a clear clinical question. Researchers should be able to explain what decision the model will improve and why existing approaches are insufficient.

2. Second, datasets should reflect the real context in which the model will be used, with representative populations and well-defined outcomes.

3. Third, models should follow sound methodological practices, including appropriate sample sizes and rigorous validation.

4. Fourth, models should be validated across multiple independent settings to ensure they generalize beyond the original dataset.

5. Fifth, researchers should study whether the model actually improves clinical decisions or patient outcomes.

6. Finally, models should be transparent and reproducible, allowing independent researchers to evaluate and improve them.

The Bigger Lesson

Prediction modeling in medicine is not primarily a computational challenge. It is a scientific and methodological challenge.

Building a model is relatively easy. Building one that is reliable, validated, and genuinely useful in practice is much harder.

As AI becomes more prominent in healthcare, the real task ahead is not simply to produce more models. It is to produce fewer, better models.