TL;DR
Researchers have developed L2G-Joint, a framework that detects anomalies in long-term medication histories by combining domain-specific medical knowledge with a "local-to-global" learning strategy. By training on synthetic but clinically informed "proxy" anomalies and focusing on the most suspicious visits rather than the whole history, the model significantly improves both the detection of fraudulent accounts and the pinpointing of specific suspicious medical visits.
What Problem Does This Paper Address?
Longitudinal healthcare records—sequences of doctor visits and prescriptions over time—are a goldmine for identifying medical errors, insurance fraud, or care-related risks. However, detecting these anomalies is notoriously difficult for three reasons:
- Sparse Signals: An anomaly might only occur in one or two visits within a years-long history. Traditional global models often "dilute" these tiny signals when looking at the entire trajectory.
- Semantic Complexity: Medication changes are normal. A patient switching from one drug to a similar one is a "plausible variation," while switching to a completely unrelated therapeutic category might be a "suspicious deviation." Generic anomaly detection algorithms can't tell the difference.
- Label Scarcity: Verified examples of medical fraud or errors are rare, making it hard to train supervised deep learning models.
Methods
The authors propose the Knowledge-Guided Local-to-Global (L2G) framework, which consists of two main innovations:
1. Knowledge-Guided Proxy Construction
Since real labels are scarce, the authors create synthetic training data using a Drug Knowledge Graph (KG) and therapeutic classification systems. They define three types of perturbations:
- K1 (Sparse Replacement): Replacing a single drug.
- K2 (Semantic Episode Replacement): Replacing a 7–30 day treatment episode.
- K3 (Segment Injection): Inserting a sequence of visits from a different patient.
Crucially, they introduce plausible negatives. For example, in K1-Same, a drug is replaced by one in the same therapeutic group (plausible). In K1-Cross, it's replaced by a drug from a distant category (suspicious). This teaches the model to find the "boundary" of normality.
2. Local-to-Global Joint Learning
The architecture processes data in stages:
- Medication-Multiset Encoder: Converts the list of drugs in a single visit into a fixed representation.
- Causal Visit-Sequence Encoder: A Transformer-based model that captures how visits relate to each other over time.
- Top-$\alpha$ Aggregation: Instead of averaging all visit scores (which dilutes signals) or taking only the maximum (which is noisy), the model aggregates the top $\alpha\%$ of the most suspicious visit scores to decide if the whole account is anomalous.
- Joint Objective: The model is trained to solve two tasks at once: identifying if an account is suspicious (global) and pinpointing which specific visits are the cause (local).
Datasets
The study uses longitudinal medication trajectories from two independent, unnamed cities:
- City-A: A larger dataset with 8,855 eligible accounts, used for primary training and evaluation (70/10/20 split).
- City-B: A smaller, more challenging cohort of 562 accounts, used to test the model's robustness across different urban environments (cross-city analysis).
Key Insights and Ideas
- Localization as a Training Signal: In many models, "localization" (explaining where the anomaly is) is a post-hoc step. Here, it is an integral part of the training process. Forcing the model to identify the specific perturbed visits actually helps it become better at identifying the overall suspicious account.
- The Power of Plausible Negatives: The "aha moment" here is that showing the model what a normal change looks like is just as important as showing it an anomaly. This prevents the model from flagging every medication adjustment as a red flag.
- Selective Aggregation: The Top-$\alpha$ strategy acknowledges that anomalies are concentrated. By ignoring the "noise" of healthy visits, the model maintains a high sensitivity to sparse deviations.
Key Results
On the City-A benchmark, L2G-Joint outperformed 12 baseline methods (including state-of-the-art models like AnomalyBERT and Deep SVDD):
- Account-level Detection: Achieved a Macro Average Precision (AP) of 0.7844, a 2.4% relative improvement over the strongest competitor.
- Anomaly Localization: Achieved a Localization Macro AP of 0.7215, a significant 9.2% improvement over previous methods.
- Robustness: In City-B, the model maintained its lead in localization, proving that the local-to-global logic generalizes well even when the underlying patient populations differ.
Limitations and Future Work
- Proxy Dependency: The model is trained on synthetic anomalies. While these are guided by medical knowledge, they may not perfectly capture the nuances of real-world sophisticated fraud.
- Hypothesis vs. Fact: The authors note that a detected "anomaly" is currently just a statistical deviation. It doesn't prove a clinical error occurred.
- Future Direction: The authors suggest moving toward active anomaly verification, where a model doesn't just flag a sequence but actively seeks out additional clinical evidence (like lab results or diagnoses) to confirm or refute its own "suspicion."
Summary updated