Abstract
A large database of brain imaging data from healthy, normal controls is useful to describe physiologic and pathologic structural changes at a population scale. In particular, these data can provide information about structural changes throughout development and aging. However, scarcity of control data as well as technical challenges during imaging acquisition has made it difficult to collect large amounts of data in a healthy pediatric population. In this study, we search the medical record at Vanderbilt University Medical Center for pediatric patients who received brain imaging, either CT or MRI, according to 7 common complaints: headache, seizure, altered level of consciousness, nausea and vomiting, dizziness, head injury, and gait abnormalities in order to find the percent of studies that demonstrated pathologic findings. Using a text-search based algorithm, we show that an average of 59.3% of MRI studies and 37.3% of CT scans are classified as normal, resulting in the production of thousands of normal images. These results suggest there is a wealth of pediatric imaging control data which can be used to create normative descriptions of development as well as to establish biomarkers for disease.
Introduction
To understand brain diseases or developmental disorders, scientists first need a "gold standard" of what a healthy brain looks like at every age. However, building a large database of healthy pediatric brain scans is notoriously difficult. Prospective studies—where healthy children are recruited specifically for research—face significant ethical and logistical barriers. Exposing healthy children to radiation (for CT) or the stress of a loud, confined space (for MRI) is difficult to justify. Furthermore, children often require sedation to stay still during an MRI, adding risk and cost. Consequently, most pediatric "control" datasets are small, making it hard to map the high variability of normal brain development.
This research proposes a shift in strategy: instead of scanning new healthy children, researchers can mine the thousands of existing scans from children who were imaged for a medical concern but whose results were perfectly normal.
Methodology
The researchers developed a pipeline to mine the Electronic Medical Records (EMR) at Vanderbilt University Medical Center using the Synthetic Derivative, a massive, de-identified database. The approach followed a specific filtering "funnel" to identify potential control images.
The pipeline involved:
- Demographic Filtering: Selecting patients under 18 years old.
- Modality Filtering: Identifying those who received a head CT or brain MRI.
- Clinical Indication Filtering: Focusing on seven common complaints: headache, seizure, altered consciousness, nausea/vomiting, dizziness, head injury (concussion), and gait abnormalities.
- Text-Search Algorithm: Searching radiology reports for "normalizing" phrases like "Normal study" or "Negative brain MRI."
- Validation: Manually inspecting a random sample of reports to verify the algorithm's accuracy.
Results and Data Analysis
The study identified a vast number of subjects across various clinical indications. For instance, over 18,000 subjects presented with headaches and over 55,000 with nausea and vomiting. The findings revealed that a significant portion of these clinical scans are interpreted as normal, representing a "hidden goldmine" of control data.
Key quantitative insights include:
- MRI Utility: Across the seven complaints, an average of 59.3% of MRI studies were classified as normal. Notably, 69.1% of MRIs for headaches and 62.4% for seizures were normal.
- CT Utility: An average of 37.3% of CT scans were normal. "Headache" CTs had the highest normality rate at 80.1%. The lower average for CTs likely reflects their use in emergency settings for acute trauma where pathology is more probable.
- Validation: Manual inspection confirmed the effectiveness of the text-search method, though the "true" rate of perfectly healthy controls after rigorous validation might be slightly lower (estimated between 35.2% and 46.4% for MRI).
Discussion
The "negative" results in a clinical setting—where a scan shows no pathology—are highly valuable for researchers as "positive" controls. By automating the search through EMRs, researchers can identify thousands of normal images without the high costs and ethical hurdles of dedicated clinical trials.
However, some limitations remain. A "Normal study" report might overlook subtle features relevant to specific research questions. Additionally, children scanned for symptoms like dizziness may not be "perfectly healthy" in the same way as recruited volunteers, introducing potential selection bias. Future work aims to refine these text-search methods to filter out minor incidental findings and ultimately create a high-resolution pediatric brain growth atlas to map maturation from infancy through adolescence.
Conclusion
There is a wealth of pediatric imaging control data available within hospital archives. By leveraging automated text-mining of radiology reports, researchers can unlock these datasets to create normative descriptions of development and establish biomarkers for disease, bypassing traditional barriers to pediatric neuroimaging research.
Summary updated