🤖 AI

Medication-Aware Financial Exploitation Detection for Alzheimer's Patients Using Edge-Aware Interaction Risk Modeling

This paper proposes a medication-aware framework that integrates medication adherence with transaction monitoring to significantly improve the detection of financial exploitation in Alzheimer's patients, demonstrating that medication context serves as a crucial modifier for identifying high-risk financial events during periods of cognitive vulnerability.

Original authors: Farzana Akter, Lisan Al Amin, Rakib Hossain, Chaitanya Gunupudi, Faisal Quader

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Farzana Akter, Lisan Al Amin, Rakib Hossain, Chaitanya Gunupudi, Faisal Quader

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine an elderly person with Alzheimer's trying to manage their finances. Sometimes, their mind is sharp, and they can spot a scammer easily. Other times, they miss a dose of their medication, and their brain feels a bit "foggy," making them much more likely to fall for a trick.

This paper proposes a new kind of "digital safety guard" that doesn't just watch the money; it also watches the medicine.

Here is how the researchers built and tested this idea, explained simply:

The Problem: The "Blind Spot"

Standard fraud detectors are like security cameras that only look at the transaction. They ask: "Is this amount weird? Is this store new?"
But they miss a crucial detail: Is the person's brain in a vulnerable state right now?
If a person is confused because they missed their medication, a transaction that looks "okay" to a computer might actually be a huge red flag to a human. The paper argues that we need to combine the "money story" with the "medicine story."

The Experiment: A Digital Sandbox

Since you can't easily test this on real people without risking their safety, the researchers built a massive digital simulation.

The Cast: They created 180 "virtual patients" living for 45 days.
The Action: These virtual people made over 30,000 fake bank transactions and took (or missed) their daily pills.
The Twist: The simulation was smart. When a virtual patient missed a pill, the system created a "vulnerability window" (a 12-hour period where they were more likely to be tricked). During these windows, the simulation made it more likely for "bad actors" to try to exploit them.

The Three "Detectives"

The researchers tested three different ways to spot the fraud to see which one worked best:

The Money-Only Detective: This detective only looks at the bank transaction. "Is this a weird amount? Is this a new store?"
The Additive Detective: This detective looks at the money plus a simple note: "Did they miss a pill today?" If yes, it adds a little extra suspicion.
The Interaction Detective (The Winner): This is the smartest one. It doesn't just add the pill info; it changes how it reads the money.
- Analogy: Imagine a stranger asking for money. If you are fully awake, you might think, "That's suspicious." But if you are half-asleep (because you missed your meds), that same stranger becomes terrifyingly dangerous. The Interaction Detective realizes that the same transaction is much riskier when the person is vulnerable.

What They Found

The results were surprising but clear:

Globally (looking at all transactions): The "Money-Only" detective was actually the most balanced overall. It didn't cry wolf too often.
During the "Danger Zones" (when meds were missed): This is where the Interaction Detective shined.
- The Money-Only detective caught about 74% of the scams happening during these vulnerable times.
- The Interaction Detective caught 91% of them.

The Trade-off: To catch those extra 17% of dangerous scams, the Interaction Detective had to flag a few more "false alarms" (transactions that were actually safe but looked suspicious). The paper argues this is a good trade: it's better to have a human check a few extra safe transactions than to miss a real scam when someone is confused.

The Big Takeaway

The paper concludes that medicine isn't a magic switch that says "Stop all spending." Instead, it's a volume knob.

When the patient takes their meds, the "risk volume" is low.
When they miss a dose, the "risk volume" turns up. The system then says, "Hey, this transaction looks a little weird, and since they missed their meds, let's treat it as a major emergency and ask a caregiver to check it."

The Bottom Line

This study didn't build a real app for banks to use tomorrow (it was a simulation). But it proved a vital concept: To protect people with Alzheimer's, we can't just watch their wallets; we have to watch their health, too. By syncing the two, we can catch more scams exactly when the person is most vulnerable, without taking away their independence when they are feeling well.

Technical Summary: Medication-Aware Financial Exploitation Detection for Alzheimer's Patients

Problem Statement
Financial exploitation poses a significant threat to individuals with Alzheimer's disease, particularly during periods of reduced cognitive stability. Conventional fraud detection systems typically rely solely on financial behavioral patterns, ignoring clinically relevant factors that dynamically alter a patient's vulnerability. Existing research often treats vulnerability as a static condition or focuses on general caregiving and communication tools, failing to integrate dynamic clinical context—specifically medication adherence—into transaction-level risk modeling. Consequently, standard anomaly detection may fail to interpret the same transaction differently based on the patient's current clinical state (e.g., whether they have missed a scheduled dose).

Methodology
The paper proposes a medication-aware framework that synchronizes medication adherence records with financial transaction streams to detect cognitively risky financial events. The methodology is structured around three layers: a perception layer for event capture, a reasoning layer for feature alignment, and an action layer for risk scoring.

Dataset Construction: A hybrid simulation dataset was generated involving 180 synthetic patients over 45 days, yielding 8,100 medication records and 30,855 transactions. Medication adherence was modeled as a daily scheduled event where a missed dose triggers a "vulnerability window" (lasting 12 hours). Transactions were generated using patient-specific Poisson processes, with burst activity and exploitation probabilities increasing during these vulnerability windows.
Feature Engineering: Five key features were extracted for each transaction:
1. Amount Anomaly ( $A_i$ ): A robust positive deviation score based on patient-specific median spending.
2. Vendor Novelty ( $V_i$ ): Binary indicator of whether the vendor is unfamiliar.
3. Frequency Spike ( $F_i$ ): Deviation in transaction count within a one-hour window.
4. Time Deviation ( $T_i$ ): Risk scoring based on the hour of the day.
5. Medication Adherence ( $M_i$ ): Binary indicator (1 for taken, 0 for missed).
Predictive Modeling: Three logistic regression formulations were evaluated:
1. Financial-Only Baseline: Uses features $(A, V, F, T)$ .
2. Additive Medication-Aware: Adds $M_i$ as a direct feature.
3. Interaction-Aware Model: Retains original features and adds interaction terms $(A(1-M), V(1-M), F(1-M), T(1-M))$ . This formulation tests whether missed medication alters the interpretation of financial anomalies.
Evaluation: Models were trained on 80% of patients and tested on 20%. Performance was measured using accuracy, precision, recall, F1-score, ROC-AUC, and Average Precision (AP). Crucially, a secondary evaluation restricted the test set to transactions occurring inside vulnerability windows to assess safety-critical performance.

Key Contributions

Framework Innovation: Introduction of a medication-aware risk modeling framework that treats adherence as a contextual modifier of financial risk rather than a static patient descriptor.
Data Synthesis: Construction of a synchronized hybrid dataset linking daily medication events with transaction streams, explicitly modeling the temporal coupling between missed doses and increased exploitation risk.
Model Comparison: A comparative analysis of interpretable weighted scoring, additive baselines, and interaction-aware models, demonstrating that interaction terms better capture the clinical hypothesis.
Contextual Evaluation: Performance assessment that distinguishes between global detection capabilities and sensitivity specifically within medication-induced vulnerability windows.

Results

Global Performance: The financial-only baseline achieved the highest global F1-score (0.5000). The interaction-aware model had a slightly lower global F1 (0.4826) but achieved the highest Average Precision (0.4476) and recall (0.6407), indicating superior ranking of high-risk cases.
Vulnerability Window Performance: The most significant finding occurred during missed-medication windows. The interaction-aware model improved recall from 0.7442 (financial-only) to 0.9070, with only a marginal decrease in F1-score (0.5865 vs. 0.5926). This suggests the model successfully identifies a higher proportion of true exploitation events when the patient is clinically vulnerable.
Feature Interpretation: Coefficient analysis revealed that vendor novelty is the dominant signal. However, the interaction terms $V(1-M)$ and $A(1-M)$ were positive, confirming that unfamiliar vendors and abnormal amounts become significantly more suspicious when medication is missed. Medication adherence itself acts as a contextual modifier rather than a standalone predictor.

Significance and Claims
The paper claims that medication adherence is most effective not as an isolated predictor of fraud, but as a contextual modifier that changes how financial anomalies should be interpreted. The study demonstrates that synchronizing clinical and financial signals allows for more context-aware protection for cognitively vulnerable individuals.

The authors emphasize that the framework is designed for edge-aware, privacy-sensitive deployment, converting raw data into compact risk features to avoid exposing sensitive clinical or banking histories. They argue for a layered protection strategy: a financial-only detector for general screening, augmented by an interaction-aware model that increases sensitivity during specific vulnerability windows. This approach supports graduated interventions (e.g., caregiver review) rather than automatic blocking, preserving patient autonomy while minimizing the risk of missed exploitation during critical periods.

The paper explicitly states that these findings are based on simulation-based validation and do not claim immediate readiness for live clinical or banking deployment. The primary contribution is the proof-of-concept that interaction-aware modeling improves vulnerability-sensitive detection compared to financial-only or additive baselines.