Exploratory data assessment for predictive modeling in the healthcare industry

Challenge 

A healthcare organization needed to predict which patients would reach a critical serum-level target — but the data stood in the way. Collected across multiple facilities with different clinical practices, the dataset was large and heterogeneous. The modeling goal was clear; the reliability of the underlying data was not. It was not clear whether patterns in the data reflected substantive clinical differences or if they were purely artefacts originating from different data collection processes. The organization knew that jumping straight into supervised modeling risked burying these issues — producing results that would be difficult to interpret, hard to trust, and impossible to act on with confidence.

Approach 

Before any predictive model was built, STAT-UP used self-organizing maps (SOMs) to uncover hidden structures in the data. High-dimensional patient data were projected onto SOMs to create a stable overview of similarities and differences across and within facilities. Dozens of maps were generated on the same underlying structure and colored by clinical measures, treatment variables, and facility-related attributes. This enabled STAT-UP’s analysts to visually trace how patterns shifted across variables — and to communicate findings without requiring stakeholders to interpret complex model outputs. The SOMs revealed systematic differences in dosage strategies, facility-specific profiles, and non-obvious correlations between clinical features.

Impact 

For the first time, decision-makers gained a clear, intuitive understanding of what their data’s actually contained — and where it fell short. Rather than reacting to isolated statistics or model diagnostics, stakeholders could directly assess where data were sufficiently comparable and where structural differences would compromise meaningful analysis. This transparency increased confidence in how the modeling scope was defined and justified critical decisions — including which facilities’ data could be reliably included and which had to be excluded. The key outcome: the organization could understand, justify, and fully trust the data foundation on which its predictive models were built. What began as uncertainty became informed confidence — and technical analysis aligned with clinical intuition.