This article is more than 5 years old. Information may no longer be current.
Treatment pathways and expert systems expected to improve patient care system-wide
Click Here to Manage Email Alerts
The art of medicine is synthesizing information about the patient gathered from current and prior visits to determine the best course of action. Data combined from history, physical examination, testing and prior interventions direct further questions, testing and treatment. This art is learned from training, experience and the published literature. As a technology leader, our UPMC Eye Center has recently developed evidence-based treatment pathways that standardize diagnosis and treatment algorithms across all satellites and providers. We are now engineering expert and artificial intelligence systems to provide individualized treatment recommendations based on continuous analysis of large databases with structure-function relationships and treatment outcomes.
In ophthalmology, decisions are the result of both data and experience-based pattern recognition. Basic knowledge from residency and fellowship training is considerably enhanced during subsequent years. Such experience-based knowledge cannot easily be tested and made available in published peer-reviewed journals, but there is an increasing body of meta-analyses, reviews and published preferred practice patterns. These guidelines ultimately direct medical decisions such as which testing to perform or which treatment to propose. Chronic, slowly progressive and variable diseases in ophthalmology, such as glaucoma, present particular challenges to prospective clinical trials. Monitoring for progression of a chronic disease takes years with potentially many visits, evolving testing protocols and treatment pathways.
We propose that collective experiences and data be used to define clinical treatment pathways by collating thousands of patient experiences. For example, by analyzing an entire hospital system’s structure and function data over time from glaucoma patients’ individual patterns of progression, risk may be established. If this is tied to a treatment outcomes analysis, lower-risk procedures can be recommended for the appropriate progression risk — for instance, office procedure or ab interno trabeculectomy vs. higher-risk trabeculectomy or tube shunt surgeries. Thankfully, with the advent of electronic databases of health records, data-mining has become a viable method of identifying and testing clinical treatment pathways.
Data-mining, machine learning and expert systems
One prominent recent story of data-mining and machine learning followed the IBM Watson project win at Jeopardy! Watson was a computer system that analyzed Jeopardy!’s answers, identified important phrases and then generated multiple proposed questions with associated measures of confidence. Applied to health care, similar technology could be used to glean relevant information from electronic records, transcriptions and other medical data. These could then be collated and ranked to answer questions such as, “How likely is it that this patient’s disease will progress if treated with drug X vs. drug Y?”
Data-mining has also been used with increased frequency and success in marketing and commerce and provides a good example of how powerful recommendations can be achieved that an individual would not be able to derive. Netflix created a recommendation system that uses thousands of subscriber ratings to predict which shows another subscriber would likely rate highly.
Another popular example of data-mining is how Target used purchasing information to determine if a customer may be pregnant and thus direct advertisements — in at least one case before the customer’s family was aware of the pregnancy. Examining the purchase histories of many women who signed up for Target’s baby registry, they were able to identify the changes in purchasing habits that foretold an impending birth. Thus, the marketers were able to target women in their second trimester before they announced to the store that they were pregnant through a baby registry.
Applications to ophthalmology
We are developing systems and methods for the UPMC Eye Center to predict how a patient would react to a particular medication or treatment given their history, past progression, other risk factors and other medications, or to identify multiple events portending future progression in eye diseases. These decision rules would then form a clinical expert system warning of potential changes in patients’ diseases.
Combining multiple techniques for data-mining and machine learning will create powerful clinical support systems. For glaucoma, thousands of clinical data points (vision, IOP, etc.) along with testing results and treatment outcomes provide valuable information to analyze retrospectively, concurrently and prospectively. Combined with information from other medical records such as family practice clinics, lab tests, genotype and family history, a treasure trove of data becomes available for each patient. A Watson-style process could extract thousands of potentially useful medical facts from disparate EHR systems, transcripts and diagnostic testing output (and recorded interpretations) for a single patient or thousands of patients. Data-mining and machine learning algorithms, such as those used by Netflix and Target, could then sift through this mountain of data to identify subtle trends and hints in the data that could be useful to predict progression of disease or efficacy of different treatments. An ophthalmologist would ultimately use these suggestions to aid in testing decisions and treatment recommendations.