“We need a cost-effective, high-quality health care system, guaranteeing health care to all of our people as a right.”
Hillary Clinton
A typical visit to the clinic remains a suboptimal experience with ample room for improvement for both the clinician and the patient/family. Many issues are related to the lack of a sophisticated digital and AI strategy before, during, and after. An “AI-enabled” clinic can improve many aspects of the entire clinic experience for both the clinician and the patient:
Before the clinic visit
Even with electronic medical records available in most clinics, the pre-visit referral and preparation processes often lack logistical efficiency. This inadequate pre-visit preparation can lead to delays in diagnosis or even an incorrect diagnosis.
Problem: New referrals are not appropriately prioritized by age or situation.
A two-week-old with respiratory distress should be seen much sooner than a teenager with episodic chest pain not associated with exercise for the past few years. There is usually no prioritization strategy for which patients need to be seen earlier than others.
Solution: Natural language processing with machine learning can prioritize new urgent patient referrals based on age and clinical condition.
Although there is a paucity of reported studies using NLP for patient referrals, a natural language tool such as clinical BERT (bidirectional pre-trained encoder representations from transformers) combined with machine learning can prioritize patients according to urgency.
Problem: Vital signs and rhythm are only episodically assessed in the clinic.
As vital signs are measured in the clinic, often these numbers do not reflect “real-world” data outside the clinic. There is increasing attention on the significance of real-world data as this data may very well be a better reflection of what the patient’s status is the majority of the time. This continual monitoring is especially relevant with telehealth.
Solution: Machine or deep learning coupled with the monitoring strategy of real-world data to offer clinicians more information for decision support.
A continuous monitoring strategy is not effective unless the analytics of the data is coupled to the data stream. In addition, an AI tool that is embedded in a monitoring tool would be more helpful in providing edge analytics needed for clinicians to determine if any initiation or escalation/weaning of therapy is necessary. Real-world data is becoming increasingly important in cardiology.
During the Clinic Visit
Clinicians regularly face the stress of a busy clinic with limited time to make the correct diagnosis and a treatment plan. AI can contribute to the reduction of this cognitive burden.
Problem: Clinically relevant information is not available on all medications.
Cardiac patients are often on multiple medications, including ones prescribed by other clinicians. The information on these medications, such as drug interactions, side effects, or particular nuances, like specific populations and responses or toxicities with these medications, are often not delineated in the record.
Solution: Machine learning can be deployed to search for these potential drug-to-drug interactions and side effects and forewarn the clinician at the time of the visit.
All implications of new medications can be delineated by machine learning in populations, especially one that is ethnically relevant, and be listed for the clinicians to render decisions on. The complications from these drug-to-drug interactions (including ones not published) can be serious but avoidable with knowledge gleaned from machine learning.
Problem: Many patients have undiagnosed rare diseases or genetic syndromes.
It is common for patients with rare diseases or genetic syndromes to have these diagnoses missed for years if not decades. The patients and families often are on medical odysseys with high levels of expenses until they are appropriately diagnosed, usually after many tests are performed.
Solution: Machine learning with a graph database can suggest possible rare diseases or genetic syndromes to the clinician at the time of the visit.
In addition, a photograph of the patient can be studied with deep learning via a convolutional neural network to detect genetic syndromes as part of the screening process.
Problem: Clinicians may not remember all the details of published protocols relevant to the patient.
Many patients have conditions that have specific follow-up plans based on published guidelines. The medical record should have a clinical “GPS” that will help guide the busy clinician to the recommended follow-up while giving the clinician flexibility in the decision-making process.
Solution: Published protocols and guidelines should be “real-time” and “real-world” and embedded in EHR for more precise patient care.
The EHR can be made more “intelligent” so guidelines are embedded and changed as real-world experience. The published guidelines are currently outdated so “living” guidelines with real-time, real-world data will be much more meaningful. The clinicians will have a much decreased cognitive load as a result of these helpful guidelines.
Next week, I will list a few more problems in the outpatient setting that can be resolved with artificial intelligence.