Pitfalls of developing medical artificial intelligence: prediction of inpatient's unexpected deterioration using the hidden implication of vital parameters
Introduction
•To demonstrate the challenges encountered in creating an artificial intelligence (AI) to predict unexpected deterioration of inpatients who stayed in the general wards, primarily focusing on the impact of data leakage on the model's performance.
Methods
•The Research Ethics Committee of Chung Shan Medical University Hospital approved the study (No. CS1-22196).
•Adult inpatients of our hospital in 2021 were included, and their values for eight vital parameters (VPs) were collected.
•The unexpected deterioration was defined by unplanned intensive care unit (ICU) admission.
•The assessment of certain VPs may be omitted for general ward inpatients, particularly for GCS and SpO2.
•These omitted parameters were recorded as missing values in the training data.
•Five-fold cross-validation was employed to verify the predictive performance of the three AI models.
•After five rounds, the predictive probabilities of a positive event in the five validation sets were merged.
•Finally, the area under the receiver operating characteristic curve (AUC) analysis was conducted to assess the merged results as the predictive performance of AI models.
Results
•A total of 27,579 inpatients with 539,245 times of VP assessments were included. About 0.5% (2,601) of these VP assessments were labeled positive events.
•Within 24 hours before an unplanned admission, GCS and SpO2 measurements frequencies increased from 0.3 to 1.1 and from 0.3 to 4.8 times per day, respectively.
•The AUCs of AI8, AI5, and AI4 for predicting unexpected deterioration within 24 hours were .914, .916, and .860, respectively.
•The pairwise comparison among the three models revealed that the significance levels were .382 between AI8 and AI5, <.0001 between AI8 and AI4, and <.0001 between AI5 and AI4.
•In this study, the AI model determined whether an unexpected deterioration occurred based on the measurement frequency of the vital parameters rather than the parameters (Table 1).
Conclusion
The AI8 and AI5 perceived the subtle and indirect hints from the change of missing values and then adjusted the predicted probability to gain a better performance in training. However, the presence of data leakage in the training data may have introduced uncertainty in the performance of AI models, even though verified by cross-validation.
