Proposal for Using AI to Assess Clinical Data Integrity and Generate Metadata: Algorithm Development and Validation.

Background: Evidence-based medicine combines scientific research, clinical expertise, and patient preferences to enhance the patient outcomes and improve health care quality. Clinical data are crucial in aligning medical decisions with evidence-based practices, whether derived from systematic research or real-world data sources. Quality assurance of clinical data, mainly through predictive quality algorithms and machine learning, is essential to mitigate risks such as misdiagnosis, inappropriate treatment, bias, and compromised patient safety. Furthermore, excellent quality of clinical data is a prerequisite for the replication of research results in order to gain insights from practice and real-world evidence.

Objective: This study aims to demonstrate the varying quality of medical data in primary clinical source systems at a maximum care university hospital and provide researchers with insights into data reliability through predictive quality algorithms using machine learning techniques.

Methods: A literature review was conducted to evaluate existing approaches to automated quality prediction. In addition, embedded in the process of integrating care data into a medical data integration center (MeDIC), metadata relevant to this clinical data was stored, considering factors such as data granularity and quality metrics. Completed patient cases with echocardiographic and laboratory findings as well as medication histories were selected from 2001 to 2023. Two authors manually reviewed the datasets and assigned a quality score for each entry, with 0 indicating unsatisfactory and 1 satisfactory quality. Since quality control was considered a binary problem, corresponding classifiers were used for the quality prediction. Logistic regression, k-nearest neighbors, a naive bayes classifier, a decision tree classifier, a random forest classifier, extreme gradient boosting (XGB), and support vector machines (SVM) were selected as machine learning algorithms. Based on preprocessing the dataset, training machine learning algorithms on echocardiographic, laboratory, and medication data, and assessing various prediction models, the most effective algorithms for quality classification were to be identified. The performance of the predictive quality algorithms was assessed based on accuracy, precision, recall, and scoring.

Results: There were 450 patient cases with complete information extracted from the MeDIC data pool. The laboratory and medication datasets had to be limited to 4000 data entries each to enable manual review; the echocardiographic datasets comprised 750 examinations. XGB demonstrated the highest performance for the echocardiographic dataset with an area under the receiver operating characteristic curve (AUC-ROC) of 84.6%. For laboratory data, SVM achieved an AUC-ROC score of 89.8%, demonstrating superior discrimination performance. Finally, regarding the medication dataset, SVM showed the most balanced performance, achieving an AUC-ROC of 65.1%, the highest of all tested models.

Conclusions: This proposal presents a template for predicting data quality and incorporating the resulting quality information into the metadata of a data integration center, a concept not previously implemented. The model was deployed for data inspection using a hybrid approach that combines the trained model with conventional inspection methods.