Specialized Portfolio

Clinical Decision Diagnostics & Patient Adherence

Decoding and predicting patient adherence failures using structural behavioral economics and neuro-computational modeling.

In modern medical systems, the most robust predictive models often fail at the last mile: human behavior. Despite optimal clinical pathways, SMS reminders, and well-designed incentives, patients frequently exhibit non-adherence, delayed screenings, or drop out of clinical trials.

By applying structural behavioral economics and neuro-computational models, we can decode and predict these adherence failures. The missing link is the structural quantification of Cognitive Friction (\(\kappa\)), Present Bias (\(\beta\)), and Trust Depreciation (\(\delta\)).

Visualizing Patient Decision Drift

Why does a patient postpone scheduling a critical screening or delay preventive care despite knowing the benefits? We model this structurally as Patient Decision Drift—the transition from optimal medical intent into a state of indefinite delay, governed by specific behavioral parameters.

%%{init: {'theme': 'neutral', 'themeVariables': { 'fontSize': '16px', 'fontFamily': 'Inter', 'edgeLabelBackground':'#ffffff', 'tertiaryColor': '#f8fafc', 'lineColor': '#475569'}, 'state': {'edgeLength': 150}}}%% stateDiagram-v2 direction LR classDef delayState fill:#fef2f2,stroke:#dc2626,stroke-width:2px,color:#991b1b,font-family:Inter classDef optimalState fill:#f0f9ff,stroke:#0369a1,stroke-width:2px,color:#0c4a6e,font-family:Inter classDef defaultState stroke-width:2px,font-family:Inter,color:#334155 [*] --> Intent : Medical Recommendation Intent --> Adherence : Proactive Action (Low β, Low κ) Intent --> Delay : Friction (κ) > Utility Intent --> Delay : Present Bias (β) Dominates state Delay { direction TB Procrastination --> CognitiveFatigue : Repeated Hurdles CognitiveFatigue --> ChronicNonAdherence : Trust (δ) Decays } Delay --> Adherence : Intervention / Friction Reduction Delay --> [*] : Loss to Follow-up Adherence --> [*] : Health Outcome Realized class Delay delayState class Adherence optimalState class Intent defaultState
Figure 1. Patient Decision Drift: Maps how individuals transition from clinical "Intent" to "Delay" when administrative friction (\(\kappa\)) outweighs utility, or when present bias (\(\beta\)) heavily discounts future health outcomes.

The Clinical Bottleneck: Intervention Fatigue

Despite the abundance of EHR data and digital health tools, patient non-adherence and clinical trial attrition remain massive challenges. Traditional behavioral economics relies on static "nudges" (e.g., text reminders, generic financial incentives), assuming uniform and stable patient responses.

However, real-world patients experience Cognitive Fatigue (\(\kappa\)) and Trust Depreciation (\(\delta\)) over time. A nudge that works on day 1 often triggers avoidance by day 14.

Optimizing Trial Retention

Predicting which cohorts require structured financial incentives versus purely informational support to remain in long-term studies.

Chronic Disease Management

Designing adaptive intervention schedules that self-adjust based on predicted fatigue levels, avoiding the "notification blind-spot."

Health Equity & Trust

Identifying vulnerable subgroups whose trust can be effectively rebuilt through targeted, low-cognitive-burden interventions.

Tech Spotlight: Local Privacy-Preserving AI

A critical challenge in applying advanced AI to clinical settings—especially in neuro-ethics, behavioral mapping, and decision science—is the privacy of Electronic Health Records (EHR). Relying on cloud-based LLM APIs introduces significant friction regarding HIPAA compliance, data sovereignty, and ethical boundaries.

To solve this, TRIBE v2 (Trust-Reinforced Intertemporal Behavioral Engine) is structurally optimized for Apple Silicon and Metal Performance Shaders (MPS) via the MLX framework.

100% Local & Offline Processing

By deploying "Local System 2" reasoning models directly on internal medical workstations, we run deep behavioral digital twin simulations completely offline. We achieve high-fidelity predictions of patient adherence while ensuring sensitive patient data never leaves the hospital's secure environment.

Case Study: Quantifying the Friction of Preventive Healthcare

My Job Market Paper serves as a foundational case study in applying this methodology to real-world clinical behavior. To understand why individuals delayed preventive healthcare (vaccination timing) despite overwhelming public health urgency, I designed a discrete choice experiment to map the intertemporal decisions of populations facing varying levels of administrative and cognitive friction.

Behavioral Parameter Recovery

By integrating these choices into a structural econometric framework, the research successfully recovered core behavioral parameters directly from empirical decision data:

  • Quantifying \(\kappa\) (Friction): Measuring exactly how much "sludge" (e.g., confusing scheduling platforms, logistical delays, wait times) it takes to deter an otherwise willing patient.
  • Isolating \(\beta\) and \(\delta\): Disentangling procrastination driven by present bias from non-adherence driven by systemic mistrust.

This approach transitions patient adherence from an unpredictable, qualitative frustration into a calculable, actionable metric for clinical trial designers, public health policymakers, and neuro-ethics researchers.

Synthetic Trial Dashboard: Adherence Simulation

To demonstrate the predictive power of the TRIBE-v2 engine, this dashboard simulates a 60-day preventive health intervention. We compare a Standard Fixed-Nudge approach (e.g., generic weekly reminders) against a Dynamic Behavioral Twin approach that anticipates cognitive fatigue (\( \kappa \)) and trust depreciation (\( \delta \)), adjusting intervention intensity in real-time.

Patient Adherence Over 60 Days

Simulated 60-day trial (n=500). TRIBE-v2 maintains 78% adherence vs 34% in standard protocol.

Intervention "Dosage" Optimization

TRIBE-v2 predicts fatigue windows (Days 20-30) and dynamically increases incentive structures.

Intervention ROI
+130%
Cost-per-retained patient
Fatigue Mitigated
82%
Algorithmically preempted
Synthetic Cohort
n=500
Monte Carlo Calibrated