This two-part series complements the Tools of the Trade podcast, where we are lucky to serve as guests for the final episode on AI across the data journey. Part 1 begins where every data pipeline does: with the sources, ingestion, and standardization work that pharmacist expertise makes reliable.
Evolving Technology, Evolving Practice Implications
Artificial intelligence continues to grow into pharmacy across the data life cycle. As this growth accelerates in breadth and depth, it still does not replace the pharmacist value in the data life cycle. The most impactful consideration is not “if” a pharmacist can approach data using AI but rather “how” a pharmacist can apply AI effectively. Clinical reasoning applied to quantitative problems is a longstanding topic across peer-reviewed literature, conference presentations, and endless meetings. Its discussions range in scope from daily operations to long-term organizational strategy. Standardizing data through FHIR and managing data volume in population health have been present for decades. The themes hold consistent while the vocabulary changes. AI does distinct work inside pharmacist supported data pipelines. It augments that work and shortens cycle time, giving pharmacists the bandwidth to apply clinical judgment more effectively rather than just implementing the same workflow at a greater velocity.
Natural Language Processing as an Early Frame for AI
Natural language processing, a subset of AI, is a clear example of how advanced computing has shaped daily clinical workflow without many in healthcare recognizing the impact. Processing SOAP notes to find the most profitable billing approach directed the architecture of structured fields in electronic health records. Applied to unstructured text, NLP translated the nuances of clinical interaction and documentation into structured, billable outputs. Today’s application and discussion of AI is recognized mainly in the form of large language models and most visible in clinical documentation through ambient scribes. Vendors now leading that space, such as Abridge, build on more than a decade of converting voice data into clinical documentation. The clinical interactions that form the bedrock of care carry a minimally invasive role here across the process of recording conversations, transcribing outcomes, and generating SOAP notes. Ubiquitous challenges and opportunities for AI in clinical settings arise with the unique structures of SOAP notes and other forms of clinical documentation can encompass structured data, unstructured data, and semi-structured data.
The same lineage as NLP points toward broader AI and LLM uses. Such use cases include scanning EHR notes for drug-drug interactions missed during medication reconciliation, assessing clinical trial eligibility against constantly growing databases, and operational functions such as communicating frequent policy change.
340B is a clear case. Its significant growth across health systems brings the challenge of ongoing policy change. If AI can better parse and process that policy, pharmacist expertise becomes important for narrowing it to local contracts. Continuous change across 340B contracting, reimbursement, and quality measures creates a rich environment where AI application and pharmacist clinical knowledge grow together, much as NLP did with the growth of electronic documentation platforms.
Data Ingestion and Standardization
Data never arrives clean. Messy data is the norm which presents a clear opportunity for pharmacists to direct AI in a pipeline rather than clean up after AI automation. The standard tasks of finding missing values, addressing erroneous observations, and deciding what to drop, align with the strength of AI to identify patterns. A patient body weight recorded as zero or negative is an obvious value that the cleaning process must address. "Weight at zero or below should be flagged" is the kind of rule that AI can replicate and scale easily.
Pharmacist expertise can enrich the data cleaning and standardization process through the data dictionary. Serving as a structured dictionary with a data set, this data dictionary establishes patterns, exposes deviations from them, and shapes the iteration of future cleaning. By positioning the data dictionary as the guide for ingestion, a pharmacist becomes the clinical expert who can harness AI as a medication expert rather than the timely and costly data analyst at the end of a poorly oriented pipeline. Setting standards across the data life cycle complements ingestion directly. Developing, applying, and iterating standards makes AI's repeating functions most useful. “AI can’t recognize what ‘good’ looks like” because AI cannot contextualize on its own. For example, when data is moved from a source system to a data warehouse in an Extract, Transformation, Load (ETL), the process of data transformation can be made routine if needed when moving from one location to another based on the schema or structures at the destination warehouse. The structure of the destination warehouse can be identified or structured by pharmacists working alongside technical experts.
More broadly, pharmacists can help structure the data dictionaries and identify which data structures matter most in the process of data cleaning. A choice as basic as identifying medications by NDC rather than RxCUI impacts the resulting analysis. That identification of drug data structure rests on the clinical expertise that set the initial standard. Pharmacists are trained to treat the patient, not the labs. In clinical practice, a lab value in “normal range” can still be an indicator of something amiss. In an automated system, a value that is clinically wrong can be read as “normal” because AI models do not understand context. In the process of Quality Checks, monitoring checks can require a human verification or validation such that a check of uniqueness, when preventing duplicate data, can align with completeness, ensuring a similar “shape” of data in the matched number of rows/columns. If left to an automated system, some quality checks may be deemed “successful” if repeated data fulfills both the “uniqueness” and "completeness" quality checks but may not reflect truly accurate data. For example, if drug identification for dispensing data is structured as a 10-digit NDC format in a source data set but must fit a 11-digit NDC format in a destination data set, then an automated system of ingestion may drop, exclude, or duplicate inaccurate data. The pharmacist's ability to read clinical inconsistency completed AI’s ability to detect an anonymous in a quantitative pattern.
Across ingestion and standardization, pharmacist expertise builds efficiency and prevents upstream AI errors.
Learn more about these topics in practice with the following Information Bytes podcast episodes:
This post was co-authored by Clay M. Cooper, PharmD and Ravi Patel, PharmD