A well-accepted truism in healthcare is that “good” patient data allows for “good” care. While defining high-quality care is complex, “good” patient data is accurate, timely, easily accessible, and relevant to the clinician’s needs at the moment.
Getting “good” data to the point-of-care is challenging for several reasons. Patient medical records are fragmented across different systems, each of which are optimized to support billing and medical liability management versus care delivery. The patient data itself is a hodgepodge of structured EMR tables, medical claims, and free-text medical notes. At best they are organized by chronology. As such, clinicians must muddle through EMRs to understand the patient, self-identify critical insights that could impact care, and identify care improvement opportunities.
To address this lack of actionable insights, many healthcare organizations employ clinicians and non-clinical staff to review charts manually. Alas, this is cost prohibitive for all but the most critical workflows.
Past implementations of chart analytics promised to automate this work of summarizing the patient, finding gaps, and identifying the “right next steps.” However, they seldom gained traction with clinicians. They were not used because they were not useful. Specifically, they were characterized by:
- Incomplete data, using only a small set of information for specific recommendations, (e.g., drug-drug interactions).
- Inaccuracy, generating false alerts and guidance
- Not conforming to the provider’s preferences and priorities
- Hard to validate and digest by the care team
Advances in Artificial Intelligence (AI) and Big Data, properly implemented, can address these challenges. Doing so will improve care delivery, patient engagement, and outcomes. Four principles to have analytics used by clinicians are:
- Patient insights must be relevant to and demanded by the care team. Simply, the alerts and patient insights must be something that providers want. This demand could stem from wanting to be more efficient, free up more time for patients, or address organizational financial needs. Further, the analytics need to account for organizational and individual variation in focus and priorities. What insights one group, or one doctors demand, may be very different than the next.
- Patient insights must be prioritized. Systemic, automated chart review can typically generate a myriad of potential improvement opportunities, especially for older and sicker patients. Because it is not practical to address every issue at once, which is time consuming for the care team and taxing on the patient, it is necessary to prioritize the recommended interventions. While patient safety and outcomes should always lead prioritization, finances and resources can also help sequence interventions.
- Patient insights should be informed by all available patient data, regardless of type or format. Combining data from different sources is hard. Tactical challenges like matching patient identity between systems is the first of many obstacles. Even within a single EMR, patient-level data like demographics and medications need to be aggregated with encounter-level information. The data that is already structured and coded, e.g., problem lists, medication lists, etc., needs to be combined with the information in the free text. For this last part, there are finally tools to extract concepts from the free text and code it appropriately. Advancements in machine learning and natural language processing (NLP) offer the potential to transform this web of information into actionable insights, driving better patient care and more efficient workflows.
- Patient insights must be accurate. As AI is employed to summarize patient charts and identify opportunities to improve care, there is more risk of getting things wrong. Without the scalability of AI, past errors of clinical analytics often come from the absence of accessible information. For example, a medication like Metformin(R) may not be ideal for a patient with low and decreasing renal function. However, if the renal function test results are buried in PDF notes, the algorithms will not incorporate kidney function into its alerting logic. With AI based chart summarization, the algorithms have a more complete data to analyze.
- The patient insights must be easy to validate. Ultimately, it is the treating provider that bares the responsibility of providing safe, effective, and compassionate care. While clinical analytics can help do this better and faster, the analytics cannot be a black box. For each recommendation, the rationale and supporting evidence needs to be easily accessible and easily digestible. The ability to prioritize relevant information and filter out less pertinent details enhances the usefulness of healthcare analytics. At KAID, we apply a “strength of evidence” score to each concept found in the notes. A high score implies a high likelihood the concept was present for a patient, a low score the opposite. For example, “MRI shows evidence of fracture” gets a high score while “fractures are common in patients with osteoporosis” gets a low score. This value allows users to get to the best evidence of a concept quickly.
Useful analytics is defined by accurate NLP models, integration of structured data, customization of search experiences, and the discernment of pertinent information. By embracing these principles, we can unlock the full potential of healthcare data, paving the way for better patient care, and more efficient processes. Clinicians will use useful tools.