In one of the modules looking at data during AIMed Pediatrics 2020, Dr. Sudhen Desai, Interventional Radiologist at Texas Children’s Hospital and Assistant Professor of Radiology at Baylor College of Medicine said paediatric medicine is not analogous to adult medicine.
Different: The heterogeneity of patient cohort
Most of the time, practitioners in the realm strive for efficiency to reduce burdens on families and young patients’ reliance on anesthesia support. For instance, in paediatric radiology, retaking images multiple times is often happening as young patients are not able to hold still. There are some potential long-term but unknown neurocognitive impact affecting them especially in the cohort of less than three years old. This can probably be omitted if artificial intelligence (AI) can come in to enhance the quality of images.
Overall, images in pediatric imaging are smaller but come with a great amount of anatomy details. Looking at an adult chest x-ray and a pediatric chest x-ray, one may find the latter covers a lot more space within the patient’s body. So, when it comes to processing the data, as Dr. Desai noted, it is important for the algorithm to be able to discern this disparity and also, the possible variances during image acquisition. Likewise, the algorithm should also account for the developmental differences existed in patients of different age groups.
He explained there are drastic differences in what considered as “normal” in plain radiography for a 2-year-old; 12-year-old, and 18-year-old patient. The algorithm has to be able to account for that. The training ground for such algorithm also needs to be complex to cater for the increase segmentation of data flow. Dr. Charitha Reddy, Clinical Assistant Professor for Pediatrics-Cardiology at Stanford University also saw a similar variability in her specialization. She said infants tend to move quickly and they have faster heart rates and a wide range of heights and weights that can affect the quality of cardiac images.
Besides, Dr. Desai touched on the importance of data governance as the lifespan of some of these pediatric patients may cover eight decades. Ensuring how data is appropriately utilized and protected will prevent discrepancy in the near or even distant future. Nevertheless, there are similar data challenges in pediatric and adult medicine, particularly in infrastructure.
Same: Rigid infrastructure are putting data in silos
Free standing hospital, like the one Dr. Desai is based right now, do not have the patient volume to drive sufficient data for the purpose of training AI models. They also lack the kind of strategic; interoperable and agile enterprise IT infrastructure that most traditional academic medical centers have. They are more focused on the functionally integrated systems like PACS (i.e., Picture Archiving and Communication System) and electronic health records (EHRs), which make it difficult to transfer huge amount of data, even within the institution itself.
Professor Neil Sebire, Chief Research Information Officer, Great Ormond Street Hospital for Children National Health Service (NHS) Foundation Trust added there was this excitement around EHRs when it was first introduced. People thought that’s where they are going to get all the data from, without realizing EHRs were designed to manage patient information. They will not provide data in the format that it suited for everything and it is a relatively small pool as compared to new data sources like home monitoring, wearables, smart hospitals that have not traditionally been in healthcare.
Dr. Reddy cited one of the approaches taken was to initiate a centralized group that will allow respective institutions to be data aggregator in which they can share and gain access to datasets that they are not individually collecting. This kind of transparent, data-sharing collaborative had been witnessed in areas focusing on with hypoplastic left heart syndrome (HLHS) or single ventricle physiology; patients after surgery, and during their time in the cardiovascular Intensive care unit (ICU).
On the other hand, Professor Sebire urged for more infrastructural investment and collaborations among various institutions to find a scalable and transferable way around standards, interoperability, definitions and data models to break through the present silos. At the same time, there is a need to have a trained professional within the healthcare system to oversee all these data curation, algorithm development, and evaluation processes.
Often, what happens is, when the data scientists come into the healthcare system, they would make requests such as “we will like to have all the lab tests for this group of patients” and they will be given thousands of rooms of raw data, rather than a nice Excel table which they have in mind. There is a missing link between translating raw data that clinicians are holding on, into what is required by fellow data scientists and engineers to do their job.
Same: Standardization and interoperability
Professor Sebire also asserted technology enabled medicine is huge and ever-extending. It encompasses many areas ranging from AI, computer vision to advanced analytics and so on. The kind of data curation and data engineering required for, let say, clinical informatics behind decision support, is going to be very different from what is required for computational biology or bioinformatics. So, there is a lot of work to be done around standardization and interoperability. This is not as easy as it seems.
For example, there were many studies being published during the COVID-19 emergency in the UK, however, it was nearly impossible to underline what data are they collecting. Even when it comes to simple thing like systolic blood pressure, it was defined differently in different studies, likewise, for ethnicity. Many of these were not dictionary definitions but created by the research teams themselves, they have also made up their own classification of data.
Dr. Peter Laussen, Executive Vice-President of Health Affairs of Boston Children’s Hospital added in physiologic data science in critical care where he specializes, definition often varies within databases. The way these terminologies are aligned will influence how information are extracted and related to care processes and decisions thereafter. Recently, there began to have roadmaps on how data should be collected, harmonized, and biases to be dealt with. At his institution, a bespoke platform was created to retrieve and permanently store data routinely collected from every critical care bed in the intensive care unit and their variability being managed.
Ultimately, Dr. Laussen hopes there will be some kind of interface between clinicians and data science, that will facilitate the transition from machine-assisted, to augmentation and eventually, automation. “Data is not only a resource but an asset. Intelligence needs data and the way information is being managed is critical for driving changes and deriving at real values in healthcare,” Dr. Laussen says.