# Clinical Context Is More Important than Data Quantity to the Performance of an Artificial Intelligence-Based Early Warning System

## TL;DR

This study shows that the clinical context of missing data is more important than the amount of data for predicting patient deterioration using AI.

## Contribution

The study introduces a novel perspective on how missing data patterns in clinical records can enhance AI predictions.

## Key findings

- Patients with high CCI scores had fewer missing data due to more testing.
- The model's performance was robust despite differences in data quantity.
- Missing data patterns were more predictive than raw data volume.

## Abstract

Background/objectives: The quantity of clinical data varies across patient populations and often reflect clinicians’ perceptions of risk and their decisions to perform certain laboratory tests. Missingness in electronic health records can be informative because it may indicate that certain clinical parameters were not measured because clinicians considered them unnecessary for stable patients. Methods: This retrospective single-center study explored the ability of a deep learning-based early warning system, the VitalCare–Major Adverse Event Score, to predict unplanned intensive care unit transfers, cardiac arrests, or death among adult inpatients 6 h in advance. We classified patients using the Charlson Comorbidity Index (CCI) and assessed whether patients with high severity and a greater volume of laboratory data benefited from more comprehensive inputs. Results: Overall, patients with high CCI scores underwent more testing and had fewer missing values, whereas those with moderate-to-low CCI scores underwent less testing and had more missing data. Within the event cohorts, however, the high-CCI and moderate/low-CCI groups showed similar proportions and patterns of missing values. The discriminative ability of the model remained robust across both groups, implying that the clinical context of missingness outweighed the raw quantity of available data. Conclusions: These findings support a nuanced view of data completeness and highlight that preserving the real-world patterns of ordering laboratory tests may enhance predictive performance.