2026-02-24
We are very good at asking how.
How to build the model.
How to tune it.
How to optimize performance.
How to extract the top features.
But in many AI-driven projects in medicine, the harder question is skipped:
Why are we building this model in the first place?
Recently, I spoke with a student who had built a machine learning model to classify individuals with a complex clinical condition using omics data.
They were proud. And they should be. Training, tuning, and validating a model takes skill.
But as they described the project, I kept returning to one question:
What problem does this solve?
This was not an isolated conversation. I have now observed multiple cases in which a model is built before the objective is clearly defined.
It reflects something deeper within our current research culture.
Technical Achievement != Scientific or Clinical Contribution
Building a working model is an accomplishment.
But accomplishment is not the same as contribution.
If your labels are derived from expert clinical evaluation, your model is trained to reproduce that evaluation.
So we must ask:
Why would someone use this model instead of a clinician?
Is it more accurate?
More accessible?
More scalable?
Does it change treatment decisions?
Does it improve patient outcomes?
If those answers are not clearly defined, then the model is a computational success, not necessarily a scientific or clinical advance.
Better Than What?
If the ground truth is derived from expert assessment, then the model is approximating that assessment.
So when we claim strong performance:
What is the comparator?
Has it been externally validated?
Does it outperform clinical variables alone?
Does it generalize across cohorts?
A model that reproduces its training labels is doing exactly what it was designed to do.
That does not automatically make it useful.
Predictive Features Are Not Therapeutic Targets
I also heard a familiar leap:
“We identified key genes from the model. Now we can design a drug.”
Machine learning identifies statistical associations that improve prediction.
Drug development requires:
Mechanistic evidence
Directionality
Biological validation
Perturbation data
A predictive feature is not automatically a disease driver.
Moving directly from feature importance to therapeutic strategy skips essential scientific steps.
The Questions We Should Ask First
Before building any ML model in a clinical domain, we should be able to answer:
What decision will this model improve?
Who will use it?
In what setting?
Why is it preferable to the current practice?
What biological hypothesis does it generate?
What is the translational pathway?
If we cannot answer these clearly, we are optimizing the technique without clarifying the purpose.
The Real Concern
This is not about talent. The students building these models are intelligent and capable.
The issue is the incentive structure.
We reward:
Novel architectures
Complex pipelines
Feature rankings
Clean performance metrics
It is entirely possible to build something technically sophisticated without ensuring it is conceptually necessary.
2026-02-17
Recently, while guiding a Ph.D. student through her project, I re-learned something important.
These moments of co-learning are deeply fulfilling as a mentor. They force you to slow down and revisit fundamentals that are easy to take for granted.
So we asked:
Who is missing what, and why?
Before running models.
Before imputing.
Before filtering features.
Missingness is often not random.
In disease datasets, missing values frequently follow patterns.
For example, a biomarker may be largely absent in severely affected patients.
If missingness differs across disease groups, that is already a signal.
If you ignore it, you risk building conclusions on biased data.
Missingness can reflect biology or process.
Data can be missing because:
The molecule is below detection limit
The sample quality differs
A batch failed
Each reason implies a different statistical approach.
Blind imputation treats all missingness the same. But it isn’t.
What pre-analysis due diligence includes:
Before modeling, examine:
Missingness per variable
Missingness per disease group
Heatmaps of missing patterns
Association between missingness and phenotype
Batch-specific dropout patterns
This is not “data cleaning.”
This is protecting your inference.
Why it matters?
If missingness correlates with disease:
Effect sizes can be distorted
Classifiers may learn data availability instead of biology
Clusters may reflect technical artifacts
At that point, you are modeling the data-generation process, not the disease.
Before asking:
“What is different between disease and control?”
Also ask:
“What is systematically unobserved between them?”
Missingness is part of the dataset’s structure.
If you treat it as noise, you may erase meaningful signal or introduce bias before you even start.