Automated Clinical Knowledge Discovery: Mining EHR Data vs. Human Intelligence

TL;DR

Building knowledge bases for Electronic Health Records (EHRs) has traditionally been a manual, labor-intensive process. This study compares Association Rule Mining (ARM)—a purely statistical approach—with Crowdsourcing—derived from real-time clinician behavior. The verdict? They are two sides of the same coin, each identifying critical clinical relationships that the other misses.

The Knowledge Bottleneck in Modern Medicine

The primary frustration for modern clinicians isn't a lack of data; it's the lack of organized data. To create an automated "Patient Summary," a system must understand that Lisinopril is related to Hypertension.

Traditionally, we relied on ontologies like SNOMED-CT or RxNorm. However, mapping local EHR data to these global standards is a nightmare of interoperability. The authors suggest a pivot: why not let the data—and the clinicians who input it—tell us what is related?

Methodology: Math vs. Collective Behavior

The researchers analyzed data from 53,108 patients, involving over 1.6 million data points. They used two distinct "engines" to discover relationships:

1. Association Rule Mining (ARM): This is "Market Basket Analysis" applied to medicine. If "Problem A" and "Medication B" appear together frequently enough (Support) and the presence of one strongly implies the other (Confidence/Chi-squared), a rule is born.
2. Crowdsourcing: This isn't Amazon Mechanical Turk; it’s "Clinical Crowdsourcing." By looking at the links clinicians manually created during the e-prescribing process, the system captures the collective intelligence of thousands of medical experts.

Note: The study used Chi-squared for ARM ranking and a Logistic Regression predictor for Crowdsourcing ranking.

Insights from the Results

The methods identified a vast number of pairs:

• ARM: 19,586 pairs
• Crowdsourcing: 31,440 pairs

Interestingly, the overlap in their "Top 500" lists was relatively small (186 pairs). This suggests that the two methods have different Inductive Biases:

• ARM's Strength: It found "rare" connections. Because it looks purely at co-occurrence, it could identify rarely prescribed drugs (like glycopyrrolate) that a clinician might not take the time to link manually.
• Crowdsourcing's Strength: It captured "Clinical Intent." For example, Metformin is primary for Diabetes, but clinicians frequently use it for Polycystic Ovarian Syndrome (PCOS). Crowdsourcing caught this secondary indication, whereas ARM might have buried it under the noise of more common associations.

Note: Spearman's rho of 0.539 indicates a moderate positive correlation, proving both methods are moving in the right direction but capturing different signals.

Critical Analysis: Why This Matters

The most significant takeaway is that neither method is sufficient alone.

• Reliance on ARM risks filling a knowledge base with "administrative noise" (e.g., linking a medication to the task of "taking medication").
• Reliance on Crowdsourcing risks missing "long-tail" clinical cases where clinicians are too busy to manually link records.

For researchers building the next generation of Clinical Decision Support (CDS), the path forward is hybridization. By using ARM to suggest potential links and Crowdsourcing to validate them (or using a weighted ensemble of both), we can build an "all-inclusive, highly accurate" knowledge base that scales without manual curation.

Conclusion

This work marks a shift from "Static Ontologies" to "Dynamic Learning Systems." As EHR data grows, the ability to automatically synthesize the relationships between problems and medications will be the difference between a cluttered database and a life-saving clinical tool.