TL;DR
The paper introduces DynAmo (Dynamic Drift Anomaly Detector), a fully unsupervised, real-time algorithm designed to detect gradual behavioral changes—known as drift anomalies—in smart home and e-Health monitoring scenarios. By combining dynamic clustering with trajectory analysis, DynAmo successfully identifies behavioral shifts as they happen without requiring labeled data or an anomaly-free initial baseline.
What Problem Does This Paper Address?
In health monitoring and telemedicine, anomaly detection algorithms typically excel at spotting point anomalies—sudden, temporary spikes in data, such as a sudden fall or a brief spike in heart rate. However, drift anomalies—gradual, long-term shifts in behavior, such as a patient slowly changing their sleep onset time or experiencing progressive cognitive decline—are much harder to detect.
Existing drift detection methods suffer from three major limitations:
- Supervision Requirements: They often rely on labeled training data, which is unrealistic given how highly personalized and context-dependent human behavior is.
- The Cold-Start Problem: They require a clean, anomaly-free "reference period" to learn what is normal, an assumption that cannot be guaranteed in real-world continuous care.
- Rigidity: They are usually tuned to specific drift types and struggle to adapt to natural, non-anomalous behavioral variations like seasonal shifts.
DynAmo addresses these gaps by operating completely unsupervised and in real-time, adapting dynamically to changing baselines.
Methods
The DynAmo framework processes continuous sensor data through a four-stage pipeline:
- Input Modelling: Raw sensor signals are pre-processed into discrete, non-overlapping behavioral events (e.g., a sleep session) characterized by features like duration, start time, and interruptions.
- Dynamic Clustering: Events are grouped within fixed time intervals $\Delta$ (e.g., daily) using an adapted version of DyClee, an incremental, density-based clustering algorithm. This tracks how clusters evolve over time.
- Trajectory Generation: Instead of analyzing every noisy data point, DynAmo extracts the centroid of the densest cluster for each time interval. This filters out daily outliers and creates a clean, continuous trend trajectory matrix ($Q$).
- Predictive Strategy (DynAmo): The algorithm employs two sliding windows: a Reference window ($R$) and a Detection window ($D$). Crucially, it introduces a look-back hyperparameter ($\lambda$) that reconstructs the historical evolution of the clusters. An ensemble of trackers and ad-hoc divergence tests then evaluates whether the distribution in the detection window has drifted significantly from the reference window. If a drift is confirmed, the reference window updates to reflect the new "normality."
Datasets
The authors evaluated DynAmo on two distinct types of datasets:
- Synthetic Benchmarks: Modified versions of the well-known ARAS, VanKasteren, and PolimiHouse smart home datasets. The authors injected synthetic drift periods at the end of normal monitoring periods to simulate the gradual onset of dementia symptoms.
- Real-World Dataset (E-Linus): Derived from the daily routines of two real-world patients suffering from senile social isolation disorders. This dataset was realistically extended to span four years and featured injected drifts affecting either sleep duration (D) or the number of sleep interruptions (I).
Key Insights and Ideas
- Filtering Noise via the "Densest Cluster": Human behavior is inherently noisy. By tracking only the centroid of the densest cluster per day, DynAmo effectively isolates the core behavioral routine from random daily fluctuations.
- The Power of Look-Back ($\lambda$): Incorporating historical context ($\lambda$) allows the model to differentiate between true behavioral drifts and benign, cyclical changes (like sleeping less during hot summer months).
- Eliminating the Clean Baseline Assumption: Because DynAmo continuously updates its reference window when a stable drift is detected, it avoids the pitfall of older models that fail if an anomaly occurs during the initial setup phase.
Key Results
The authors compared DynAmo against a wide array of baselines and state-of-the-art (SoA) drift detectors using average F1 scores over 30 runs:
- Superior Performance: DynAmo emerged as the top-performing method overall, outperforming the supervised state-of-the-art method, D3, by an average of 7.91% across all datasets.
- Robustness in Complex Environments: While supervised methods performed well on highly regular synthetic data, their performance collapsed on the noisy, oscillatory real-world E-Linus datasets. DynAmo maintained stable, high F1 scores across both scenarios (e.g., achieving an F1 of 0.7699 on the challenging ELP2-D dataset).
- Efficiency: DynAmo requires only a small fraction of data (~1% of total monitoring days) to populate its windows and begin making accurate real-time predictions, effectively mitigating the cold-start problem.
Limitations and Future Work
Despite its strong performance, the authors note a few limitations:
- Lack of Unified Detection: DynAmo is specialized for drift anomalies; it does not currently integrate point anomaly detection into the same predictive framework (though point anomalies can be flagged as low-density cluster outliers).
- Explainability: The model lacks an explicit explainability mechanism to tell clinicians why a drift occurred or what specific causal factors triggered the alert.
- Recurrent Anomalies: While the paper outlines a theoretical approach to handling recurring seasonal anomalies using a behavioral history queue, fully implementing and testing this capability is left for future work.
Summary updated