Mapping drug-drug interaction risk in an 84-year-old polypharmacy patient — from medication reconciliation to network-guided deprescribing
Polypharmacy — conventionally defined as concurrent use of ≥5 medications — affects an estimated 40–50% of adults over 65 and >60% of nursing-home residents. Before any interaction can be modeled, the clinical team must first establish ground truth: what is this patient actually taking? Medication reconciliation is a deceptively difficult data-collection problem, since prescriptions are frequently scattered across multiple prescribers, pharmacies, and self-directed over-the-counter (OTC) use that never appears in the electronic health record.
A brown-bag medication review — patients bring every pill bottle, blister pack, and supplement container to the visit — remains the gold-standard reconciliation method, because EHR-derived active medication lists are wrong in an estimated 30–50% of elderly patients due to:
• Prescribing silos: a cardiologist prescribes amiodarone without visibility into the psychiatrist's sertraline order • Pharmacy fragmentation: mail-order pharmacy for maintenance drugs, retail pharmacy for acute prescriptions, so no single dispensing record captures the full list • OTC blind spots: diphenhydramine ("PM" sleep aids), NSAIDs, and herbal supplements are rarely volunteered unless specifically asked • Discontinuation drift: medications stopped by one provider are not always removed from another provider's active list ("phantom" prescriptions) • Adherence variability: pillboxes, dosing diaries, and pharmacy refill timing (proportion of days covered, PDC) help distinguish prescribed from actually-consumed regimens
For this case — an 84-year-old with atrial fibrillation, ischemic heart disease, type 2 diabetes, major depressive disorder, mild cognitive impairment, and chronic low back pain — the reconciled 11-drug list is: warfarin, amiodarone, metoprolol, simvastatin, omeprazole, sertraline, tramadol, furosemide, metformin, donepezil, and OTC diphenhydramine.
Two complementary screening instruments are run over the reconciled list before any pairwise interaction analysis begins:
• American Geriatrics Society (AGS) Beers Criteria (2023 update): an explicit list of medications generally considered inappropriate in older adults independent of interactions — flags diphenhydramine (strong anticholinergic, delirium/fall risk), tramadol (seizure risk, hyponatremia, serotonergic), and long-duration PPI use beyond 8 weeks without re-evaluation • STOPP/START (Screening Tool of Older Persons' Prescriptions / Screening Tool to Alert to Right Treatment), version 3: STOPP identifies potentially inappropriate prescriptions by physiologic system; START identifies indicated therapies that may be missing • PRISCUS list (Germany) and FORTA (Fit fOR The Aged) classification serve analogous roles in other health systems
These tools flag 4 of the 11 medications for closer review at intake, before the network-graph interaction analysis in Stages 2–3 quantifies exactly how these medications interact with the rest of the regimen.
With 11 confirmed medications, the number of possible pairwise combinations is C(11,2) = 55. Each pair is algorithmically screened against curated interaction databases, and every flagged pair is annotated with a severity tier and a mechanistic explanation — almost always rooted in shared cytochrome P450 (CYP) metabolism, additive pharmacodynamic effects, or renal/hepatic clearance competition.
Three reference sources are triangulated for this case, since no single database has perfect coverage:
• DDInter 2.0: an open-access, literature-curated database of >100,000 drug pairs classified into major / moderate / minor severity with mechanistic tags • DrugBank: structured drug-target and drug-metabolism data used to computationally infer shared-enzyme interactions not yet described in the clinical literature • Lexicomp / Micromedex: the clinical decision-support standard embedded in most EHRs, weighting real-world case reports and FDA label warnings
Severity assignment follows a consistent rubric across these sources: • Major (contraindicated / avoid combination): documented serious harm — bleeding, arrhythmia, serotonin syndrome, rhabdomyolysis — with limited safer alternatives • Moderate (monitor / dose-adjust): clinically significant effect but manageable with lab monitoring, dose reduction, or spaced administration • Minor: statistically detectable interaction with low absolute clinical impact in most patients
For this patient, screening returns 4 major interactions: warfarin–amiodarone (CYP2C9 inhibition, INR elevation), amiodarone–simvastatin (CYP3A4 inhibition, myopathy/rhabdomyolysis risk), sertraline–tramadol (additive serotonergic activity, serotonin syndrome), and donepezil–diphenhydramine (pharmacodynamic opposition — a cholinesterase inhibitor paired with a potent anticholinergic, worsening cognition and undermining the intended therapeutic effect of both drugs).
Most of the clinically significant interactions in this regimen trace back to a small set of shared hepatic cytochrome P450 enzymes:
• CYP2C9: metabolizes S-warfarin (the more potent enantiomer); amiodarone is a potent CYP2C9 inhibitor, reducing warfarin clearance by up to 30–50% and typically requiring a 30–50% empiric dose reduction plus tighter INR monitoring • CYP3A4: metabolizes simvastatin; amiodarone inhibits CYP3A4, raising simvastatin exposure — FDA label restricts simvastatin to a maximum 20mg/day when co-administered with amiodarone due to rhabdomyolysis risk • CYP2D6: metabolizes tramadol's activation to its active metabolite (O-desmethyltramadol) and metoprolol clearance; genetic polymorphism at this locus (poor vs. ultrarapid metabolizer phenotypes) substantially changes both efficacy and toxicity risk • CYP2C19: metabolizes omeprazole and modulates warfarin sensitivity indirectly through vitamin K cycle interactions
Because warfarin, amiodarone, simvastatin and tramadol collectively span all four of these enzymes, this regimen exemplifies why polypharmacy risk grows combinatorially, not linearly, with medication count — a phenomenon formalized in the network-graph analysis of Stage 3.
| Product | Indication | Trial Design | Key Result |
|---|---|---|---|
| Warfarin + Amiodarone | CYP2C9 inhibition | ↓ S-warfarin clearance → supratherapeutic INR, major bleed risk | Major — dose-reduce warfarin 30–50%, weekly INR |
| Amiodarone + Simvastatin | CYP3A4 inhibition | ↑ statin exposure → myopathy / rhabdomyolysis | Major — cap simvastatin at 20mg/day |
| Sertraline + Tramadol | Serotonergic synergy | Additive serotonin reuptake inhibition | Major — serotonin syndrome risk, avoid combo |
| Donepezil + Diphenhydramine | Pharmacodynamic opposition | Anticholinergic blocks cholinesterase-inhibitor benefit | Major — worsens cognition, delirium risk |
A flat list of 13 interaction pairs is clinically hard to reason about. Recasting the regimen as a graph — 11 nodes (medications), 13 weighted edges (interactions) — makes the risk structure visually and mathematically explicit. Graph-theoretic metrics such as degree centrality and betweenness identify which single medication, if removed, would most reduce total interaction burden — turning deprescribing into an optimization problem rather than guesswork.
The interaction network is built as an undirected, edge-weighted graph G = (V, E):
• V (vertices): 11 medications, each carrying attributes — drug class, primary metabolic pathway, renal vs. hepatic clearance, Beers Criteria flag • E (edges): 13 interactions, each weighted by severity (major=2, moderate=1, minor=0.5) and annotated with mechanism • Layout: force-directed placement (Fruchterman-Reingold-style spring model) — nodes with more/heavier edges are pulled toward the graph center; isolated or lightly-connected nodes drift to the periphery
This is directly analogous to how hospital pharmacy informatics teams and clinical decision-support vendors (e.g., embedded DDI graph modules in Epic, Cerner) visualize regimen risk for pharmacist review — the same underlying formalism used in social-network and protein-interaction analysis, repurposed for medication safety.
Three centrality measures are computed over the graph:
• Degree centrality: number of direct interaction edges per node. Warfarin has degree 6 (interacts with amiodarone, sertraline, tramadol, furosemide, omeprazole, simvastatin) — the single most interaction-prone drug in the regimen, consistent with its notoriously narrow therapeutic index and near-total dependence on CYP2C9/VKORC1 metabolism • Betweenness centrality: how often a node lies on the shortest path between other node pairs — warfarin scores 0.41, meaning it is not just heavily connected but structurally central to how risk propagates across the whole regimen • Network density: total edges divided by all possible edges — 13/55 = 0.236, meaning roughly 1 in 4 medication pairs in this regimen has a documented interaction, well above the ~0.10–0.15 density typically seen in regimens of 5 or fewer drugs
This quantifies a well-established clinical epidemiology finding: interaction risk scales combinatorially with medication count. Adding an 11th medication to a 10-drug regimen does not add one new risk — it potentially adds up to 10 new pairwise risks, which is precisely why polypharmacy prevalence correlates so strongly with adverse drug event (ADE) rates in ambulatory elderly populations (Ann Intern Med cohort data: each additional medication increases ADE odds by ~7–10%).
A generic interaction graph treats every patient identically, but real risk depends on individual physiology: renal clearance, hepatic enzyme phenotype, cardiac conduction baseline, and cognitive reserve. Stage 4 overlays patient-specific pharmacokinetic (PK) and pharmacodynamic (PD) parameters onto the network, converting qualitative severity tiers into a single composite, patient-specific ADE Risk Index.
Estimated glomerular filtration rate (eGFR) is calculated via the 2021 CKD-EPI creatinine equation (race-free version), yielding 42 mL/min/1.73m² for this patient — CKD stage 3b. This single number reshapes dosing for at least three of the eleven medications:
• Metformin: contraindicated below eGFR 30, dose-reduced (max 1000mg/day) between 30–45 — this patient sits just above that threshold, requiring close monitoring for lactic acidosis risk, especially if renal function declines further • Furosemide: reduced renal clearance blunts diuretic response, often requiring dose escalation, which paradoxically increases electrolyte-disturbance risk (hypokalemia, which itself potentiates digoxin-class and QT-prolonging drug toxicity) • Warfarin: while not renally cleared itself, reduced renal function increases sensitivity to bleeding complications and is an independent risk factor incorporated into bleeding-risk scores (HAS-BLED)
Renal function is not static — it is itself modeled as a slider parameter in this simulation because AKI episodes (common in hospitalized elderly patients, triggered by contrast dye, NSAIDs, or dehydration) can acutely shift every renally-cleared drug's effective exposure within 24–48 hours.
Three additional composite risk scores are layered onto the network:
• Anticholinergic Cognitive Burden (ACB) scale (Boustani et al.): assigns 0–3 points per medication based on anticholinergic potency. Diphenhydramine (3 points, high potency) and donepezil's pharmacodynamic antagonism combine with sertraline's mild anticholinergic activity (1 point) to reach a cumulative ACB score of 6 — well above the ≥3 threshold associated with measurable cognitive decline and 50% increased dementia risk in longitudinal cohorts (JAMA Intern Med, Gray et al. 2015)
• QTc prolongation: amiodarone alone prolongs QTc; combined with metoprolol's bradycardic effect and electrolyte shifts from furosemide (hypokalemia lowers the arrhythmia threshold), Bazett-corrected QTc is estimated at 478ms — above the 470ms threshold flagged as high-risk for torsades de pointes in women
• CYP2D6/2C9 phenotype: genotype-guided dosing (CPIC guidelines) matters enormously here — a CYP2D6 poor metabolizer accumulates active tramadol metabolite more slowly but is paradoxically more sensitive to metoprolol accumulation, while a CYP2C9 poor metabolizer combined with amiodarone co-therapy can push warfarin sensitivity into a dangerously narrow dosing window, consistent with the elevated INR projection of 4.1 modeled here.
Composite ADE Risk Index rises from 71/100 (structural network risk alone, Stage 3) to 84/100 once this patient's actual renal function, CYP phenotype, and cardiac conduction baseline are incorporated — demonstrating that network topology alone systematically underestimates real-world risk until it is personalized with PK/PD data.
The clinical payoff of building this network is deprescribing: systematically identifying which medications can be reduced, switched, or stopped to collapse the highest-risk edges while preserving therapeutic benefit. A pharmacist-led medication review, structured around STOPP/START criteria and this patient's interaction graph, targets the three medications whose removal eliminates the most high-severity edges at the lowest clinical opportunity cost.
Rather than reviewing all 11 medications with equal scrutiny, the network highlights the highest-leverage targets — nodes and edges whose removal collapses the most risk for the least therapeutic sacrifice:
• Diphenhydramine (OTC): touches 3 edges (donepezil-opposition, tramadol-sedation, sertraline-additive CNS depression) and carries a Beers Criteria flag; has a low-risk substitute (non-pharmacologic sleep hygiene, or short-term low-dose trazodone if needed) — highest-leverage, lowest-cost removal • Tramadol: touches 3 edges including the major sertraline serotonin-syndrome interaction; substituted with scheduled acetaminophen plus topical NSAID for chronic low back pain, avoiding both the serotonergic and CYP2D6-dependent risks • Omeprazole: touches 2 edges via CYP2C19-mediated warfarin sensitivity variability; after 18 months of continuous use with no re-evaluation, STOPP criteria recommend step-down or discontinuation trial given resolved GERD symptoms
Warfarin and amiodarone are retained — both are pharmacodynamically essential for this patient's atrial fibrillation and cannot be safely substituted — but the remaining major interaction (warfarin–amiodarone) shifts from "unmanaged" to "actively monitored" with a formal INR monitoring schedule increased to weekly for 4 weeks after any dose change.
After deprescribing, the network is rebuilt: 8 remaining medications, 7 remaining edges (down from 13), and network density falls from 0.24 to 0.25 of a smaller graph — but critically, major-severity edge count drops from 4 to 2, and both remaining major interactions (warfarin–amiodarone, amiodarone–simvastatin) are now under active monitoring protocols (weekly INR; creatine kinase check if myalgia develops).
The composite ADE Risk Index — combining graph density, severity-weighted edge sum, ACB score, QTc, and renal-adjusted clearance — falls from 84/100 to 34/100, a modeled 59% relative reduction. This is consistent with published deprescribing-intervention literature: the OPTIMIZE and SENATOR randomized trials of structured geriatric medication review report 30–40% reductions in adverse drug events and comparable reductions in fall-related hospitalization at 12-month follow-up when a pharmacist-led review is paired with explicit interaction-network or clinical decision-support tools, versus usual care.
The network graph does not replace clinical judgment — it prioritizes it. By quantifying exactly which medication removal collapses the most edges, a time-limited geriatrics or pharmacy consultation can focus on the 3 highest-leverage changes instead of attempting an exhaustive, unstructured review of all 55 possible pairs — the single biggest barrier cited by clinicians for why deprescribing so often does not happen in a standard 15-minute visit.