AI model performance is more influenced by missing data reflecting clinical judgment than by the sheer volume of data

AITRICS (CEO Kwang Joon Kim), a company specializing in artificial intelligence(AI) technology announced on July 2 that its research paper, which explores how missing clinical data can reflect medical decision-making rather than merely indicating information gaps, has been published in the SCIE-indexed international journal Journal of Clinical Medicine.

The study retrospectively analyzed clinical data from 24,359 adult patients admitted to general internal medicine and surgical wards at Jesus Hospital. It validated the concept of Informative Presence — the idea that missing data itself can represent the outcome of clinical judgment.

Building on previous research, the team hypothesized that data collection patterns and missing rates would vary based on patient severity. Patients were stratified into high-risk (Charlson Comorbidity Index, CCI > 3) and low-to-moderate-risk (CCI ≤ 3) groups. The study then compared missing data rates and AI model performance between these two cohorts.

The results showed that high-risk patients underwent more tests and thus had lower missing data rates. Conversely, the low-to-moderate-risk group had higher missing rates due to less frequent testing. However, within both groups, patients who experienced clinical deterioration events had consistently lower missing rates than those who did not. This suggests that clinicians tend to order more tests when they suspect potential deterioration, regardless of the patient’s baseline risk level.

Despite differences in test frequency and missing data rates across patient groups, the AI model's predictive performance remained stable — with AUROC scores of 0.86 for the overall cohort, 0.86 for the high-risk group, and 0.85 for the low-to-moderate-risk group. This indicates that the clinical context behind whether a test was performed plays a more critical role in model performance than the amount of data alone.

CEO Kwang Joon Kim of AITRICS stated, “Since test frequencies and missing patterns vary depending on patient condition, AI models must be able to recognize and interpret such clinical behavior patterns based on severity. Rather than relying solely on numerical data, models that incorporate clinical judgment will be more trusted and effectively applied in real-world medical settings.”