Electronic Health Record Mining

Identifies and classifies key medical concepts from unstructured clinical narratives. Modern systems extract medications, diagnoses, procedures, and lab values with high precision.

• Maps extracted entities to standardized ontologies like SNOMED CT and RxNorm
• Handles clinical abbreviations (e.g., 'MI' for myocardial infarction) and negation detection
• Achieves F1 scores exceeding 0.90 on benchmark datasets like i2b2
• Example: Extracting all mentions of 'metformin' and 'type 2 diabetes' from millions of discharge summaries

Reconstructs the chronological sequence of clinical events from narrative text to establish disease progression timelines. This capability is critical for identifying drug exposure windows and subsequent outcomes.

• Links clinical events to temporal expressions using TIMEX3 normalization
• Models relationships like 'before', 'after', and 'overlap' between medical events
• Enables detection of adverse drug events by analyzing what occurred after medication administration
• Example: Determining if a rash appeared within 2 weeks of starting a new ACE inhibitor

Defines and retrieves patient cohorts based on complex clinical criteria that combine structured codes (ICD-10, CPT) with unstructured text evidence. This hybrid approach dramatically improves cohort sensitivity.

• Implements computable phenotype algorithms using tools like the PheKB catalog
• Reduces false negatives by 30-50% compared to relying solely on billing codes
• Example: Identifying all patients with 'drug-resistant hypertension' by combining diagnosis codes with blood pressure values and medication mentions in clinical notes

Automatically removes protected health information (PHI) from clinical text to enable secondary research use while maintaining HIPAA compliance. Modern transformer-based models outperform older rule-based systems.

• Detects 18 HIPAA-defined identifiers including names, dates, and geographic subdivisions
• Uses contextual understanding to distinguish between patient names and medical terms
• Achieves recall above 0.98 on the i2b2 de-identification challenge benchmark
• Example: Replacing 'John Smith presented on 01/15/2023' with '[PATIENT] presented on [DATE]'

Converts free-text clinical findings into structured, queryable formats by mapping extracted concepts to standard terminologies. This bridges the gap between narrative medicine and computational analysis.

• Normalizes drug names to RxNorm concept unique identifiers (CUIs)
• Maps disease mentions to ICD-10-CM and SNOMED CT codes
• Resolves synonyms (e.g., 'high blood pressure' → SNOMED:38341003)
• Enables large-scale pharmacovigilance queries across heterogeneous EHR systems

Applies statistical and machine learning methods to mined EHR data to identify drug-disease associations and potential repurposing signals. This transforms routine clinical data into a hypothesis-generating engine.

• Uses self-controlled case series designs to minimize confounding
• Applies disproportionality analysis adapted from spontaneous reporting systems
• Detects unexpected therapeutic benefits (e.g., a diabetes drug associated with reduced cancer incidence)
• Example: Identifying metformin's potential protective effect against certain cancers from longitudinal patient records

Large-scale retrospective analysis of diabetic patient cohorts revealed that metformin, a first-line type 2 diabetes drug, was associated with a statistically significant reduction in cancer-specific mortality and incidence across multiple tumor types.

• Signal: Observational studies using structured EHR data (HbA1c, diagnosis codes) showed a dose-response relationship.
• Mechanism: Subsequent research implicated AMPK/mTOR pathway inhibition as the underlying anti-neoplastic mechanism.
• Current Status: The signal has progressed to numerous randomized controlled trials evaluating metformin as an adjuvant cancer therapy.

While sildenafil's repurposing from angina to erectile dysfunction is famous, EHR mining of pediatric intensive care units later identified its profound effect on pulmonary arterial hypertension (PAH).

• Discovery Path: Mining unstructured clinical notes and hemodynamic monitoring data revealed improved cardiac output in ventilated neonates receiving sildenafil for persistent pulmonary hypertension.
• Pharmacology: The drug inhibits phosphodiesterase-5 (PDE5) , increasing cyclic GMP and causing pulmonary vasodilation.
• Outcome: This led to the FDA approval of intravenous sildenafil (Revatio) specifically for PAH, a distinct indication from its original use.

A serendipitous clinical observation, later validated by systematic EHR review, showed that the non-selective beta-blocker propranolol caused rapid regression of infantile hemangiomas.

• EHR Signal: Analysis of medication administration records and dermatology consult notes confirmed that hemangioma ulceration and size decreased dramatically within days of propranolol initiation for unrelated cardiac issues.
• Triple Mechanism: The effect is mediated by vasoconstriction, inhibition of angiogenesis via VEGF downregulation, and induction of endothelial cell apoptosis.
• Impact: Propranolol (Hemangeol) became the first-line systemic treatment, replacing corticosteroids and drastically reducing the need for surgical intervention.

Despite its tragic history with teratogenicity, structured mining of oncology EHR data resurrected thalidomide as a cornerstone therapy for multiple myeloma.

• Data-Driven Insight: Analysis of bone marrow biopsy reports and survival data in relapsed/refractory myeloma patients showed unprecedented response rates to thalidomide monotherapy.
• Anti-Angiogenic Mechanism: The drug inhibits tumor necrosis factor-alpha (TNF-α) and blocks the formation of new blood vessels feeding the malignant plasma cells in the bone marrow microenvironment.
• Legacy: This discovery validated anti-angiogenesis as a therapeutic strategy in hematologic malignancies and spawned the development of more potent immunomodulatory imide drugs (IMiDs) like lenalidomide.

Longitudinal mining of EHR prescription data and colonoscopy/pathology reports across large health systems solidified the role of aspirin in colorectal cancer chemoprevention.

• EHR Evidence: Analysis of millions of patient-years revealed that consistent, low-dose aspirin use was associated with a reduced incidence of colorectal adenomas and carcinomas, particularly in patients with Lynch syndrome.
• Molecular Basis: The protective effect is linked to the irreversible inhibition of cyclooxygenase-2 (COX-2) , reducing prostaglandin E2 and downstream pro-carcinogenic inflammation.
• Guideline Impact: The United States Preventive Services Task Force (USPSTF) now includes aspirin use recommendations for primary prevention of colorectal cancer in specific age groups based on cardiovascular risk profiles.

Mining of anesthesia records and subsequent psychiatric encounter notes revealed that ketamine, a dissociative anesthetic, produced rapid and sustained antidepressant effects in patients with treatment-resistant depression (TRD).

• Clinical Signal: EHR analysis showed that patients receiving ketamine for procedural sedation had significantly lower depression severity scores in subsequent mental health visits compared to those receiving other anesthetics.
• Novel Mechanism: Unlike traditional monoaminergic antidepressants, ketamine acts as an NMDA receptor antagonist, triggering a rapid glutamate surge and synaptic plasticity restoration via BDNF-mTOR signaling.
• Translation: This EHR-driven insight directly led to the FDA approval of esketamine (Spravato) nasal spray, representing the first novel mechanism for depression treatment in decades.

Clinical evidence derived from the analysis of real-world data (RWD) , such as electronic health records, insurance claims, and patient registries. Unlike controlled clinical trials, RWE captures the actual usage, safety, and effectiveness of medical products in heterogeneous patient populations.

• Key Sources: EHRs, claims databases, patient-generated data
• Regulatory Impact: The FDA and EMA increasingly accept RWE to support label expansions for existing drugs
• Analytical Challenge: Requires rigorous methods to control for confounding, bias, and missing data inherent in observational datasets

The science and activities relating to the detection, assessment, understanding, and prevention of adverse effects or any other drug-related problem. EHR mining serves as a critical passive surveillance system for pharmacovigilance.

• Signal Detection: Algorithms scan clinical notes and structured data for disproportionate reporting of unexpected adverse events
• Disproportionality Analysis: Statistical methods like the Proportional Reporting Ratio (PRR) identify drug-event pairs occurring more frequently than expected
• Inverse Signals: Unexpected absence of a condition in patients taking a specific drug can flag a protective effect, creating a repurposing hypothesis

A computational approach that compares a disease's gene expression pattern against a reference database of drug-induced gene expression changes. The goal is to identify compounds that reverse the disease state at the molecular level.

• Connectivity Map (CMap): The foundational resource from the Broad Institute containing genome-wide transcriptional data from cultured human cells treated with perturbagens
• Mechanism: If a drug's expression signature is the mirror image of a disease signature, the drug is predicted to be therapeutic
• EHR Integration: Patient-level transcriptomic data extracted from biobanks linked to EHRs enables direct clinical validation of CMap predictions

A machine learning technique that projects the entities and relations of a biomedical knowledge graph into a low-dimensional vector space. This enables the prediction of missing links, such as novel drug-disease associations.

• Translational Models: Algorithms like TransE model relationships as translations in the embedding space (e.g., Drug + Treats ≈ Disease)
• Heterogeneous Graphs: Integrate diverse node types including drugs, diseases, genes, pathways, and side effects extracted from EHRs and literature
• Link Prediction: The primary task—scoring the plausibility of a treats edge between an existing drug and a new disease indication

A statistical framework that goes beyond correlation to determine whether a specific drug exposure directly causes a particular therapeutic outcome. Essential for distinguishing true repurposing signals from confounding in observational EHR data.

• Mendelian Randomization: Uses genetic variants as instrumental variables to mimic a randomized trial, assessing if a drug target's modulation causally impacts disease risk
• Propensity Score Matching: Balances treated and untreated patient cohorts on observed covariates to reduce selection bias
• Target Trial Emulation: Explicitly designs an observational study to mimic a hypothetical randomized controlled trial, specifying eligibility, treatment assignment, and follow-up

A critical validation error where information from the test set inadvertently influences the training process, leading to over-optimistic performance estimates. In EHR-based drug repurposing, this is a pervasive and subtle risk.

• Patient-Level Leakage: Multiple records from the same patient appearing in both training and test splits
• Temporal Leakage: Using future information to predict past events, such as including a diagnosis made after the index date in the feature set
• Concept Leakage: Including features that are direct proxies for the outcome, like using a drug's prescription code to predict the disease it treats
• Mitigation: Strict temporal splitting by patient index date and careful feature exclusions