When the Data Is Thin, Skewed, or Missing | AI Learning Guide

Look at the three columns and notice that they are not three degrees of the same problem — they are three different ways the evidence can mislead. Missing data forces the model to substitute an assumption for a measurement, and those assumptions cluster wherever the gaps cluster. Inconsistent data means the same property was measured differently in different places, so the model may be learning the habits of the labs rather than the behavior of the drugs. Bias is the most serious: if the historical record overrepresents certain drug classes or patient populations, the model applies patterns from that slice to everything, including drugs it was never built to describe. The conclusion underneath all three is the one worth carrying forward. A confident-looking score is not automatically stronger evidence than a cautious one. What matters is what the model was fed.

0:00 / 0:00

Auto-play

Three ways data quality limits a prediction

Missing

Partial records: a structure but no assay results, or an outcome with no consistent safety data
The model fills gaps with assumptions, not measurements
Blind spots concentrate wherever the gaps concentrate

Inconsistent

The same property measured by different labs, instruments, and formats
The model reconciles disagreements that may not be real
The learned pattern can reflect how data was collected, not the drug

Biased

The record overrepresents certain drug classes, diseases, or populations
Patterns from that slice are applied to everything
Underrepresented drugs are judged by patterns never built for them

Data quality does not just reduce accuracy — it changes what a prediction is entitled to claim. A confident score built on a thin or skewed record is not stronger evidence than a cautious score built on a rich one; it is a claim that has not been tested against its own blind spots.