[October 7, 2026] AITRICS, a medical artificial intelligence (AI) company, announced on October 7 that its research on MADE (Masked-Attention-based timestamp-wise Data Embedding), a patient deterioration prediction AI model, has been published in the Journal of the American Medical Informatics Association (JAMIA), an internationally renowned journal in the field of medical informatics.

The paper, titled "MADE: leveraging informative missingness and temporal evaluation for electronic health record-based deterioration prediction in general wards," presents a new approach that uses missingness in electronic health record (EHR) data to continuously predict deterioration in general ward patients.

EHR data is collected irregularly, because which variables are measured, and when, depends on the patient's condition and clinical judgment. As a result, some variables are left unmeasured. Existing EHR-based AI models have addressed these missing values through imputation, such as carrying forward previously observed values or replacing them with statistically estimated ones.

In this study, the AITRICS research team focused on the possibility that the very fact that a variable was not measured may itself carry information about a patient's condition. The team used a masked-attention Transformer and a bidirectional long short-term memory network (bi-LSTM) to analyze observed values and missingness patterns together. The model predicts deterioration risk using not only the values observed at each timestamp but also the pattern of which variables were not measured.

MADE was evaluated on an internal cohort from Severance Hospital (SVH) and two external cohorts from Asan Medical Center (AMC) and Hallym University Medical Center (HUMC). The team compared it with conventional and deep learning-based models in predicting major adverse events (MAE) and sepsis. In ablation studies, it also compared a configuration with no imputation against existing imputation methods, namely last observation carried forward (LOCF) and multiple imputation by chained equations (MICE).

The findings suggest that missing information in EHR data can be used as potentially informative evidence about a patient's condition, rather than treated merely as a gap to be filled. External validation using data from different medical institutions also assessed MADE's performance across varying data collection environments and characteristics.

MADE is not confined to academic research. It is applied to AITRICS-VC (VitalCare), the company's AI solution for predicting patient deterioration, and is in operation in real clinical settings. By presenting research on AI technology already in clinical use to an international journal, AITRICS plans to continue expanding research on, and clinical application of, EHR-based patient deterioration prediction.

Yechan Mun, Researcher at AITRICS, said, "This study suggests that missingness in EHR data, which inevitably arises in clinical practice, can be used as informative evidence for understanding a patient's condition, rather than viewed merely as something to be filled in. Moving forward, AITRICS will continue to research, develop, and validate medical AI technologies that can be used in real clinical settings."