🤖 machine learning

Agentic Anomaly Detection with ORCA-Style Dynamic Inductive Bias Adaptation in Multimodal Wearable Time Series Data

The paper introduces ORCA, an agentically controlled framework for wireless body area networks that dynamically adapts temporal receptive fields at inference time based on lightweight signal statistics, achieving robust anomaly detection performance across heterogeneous physiological data without requiring retraining or dataset-specific tuning.

Original authors: Anushka Roy, Jyotirmoy Singh, Shreea Bose, Chittaranjan Hota

Published 2026-08-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Anushka Roy, Jyotirmoy Singh, Shreea Bose, Chittaranjan Hota

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ 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 you are wearing a high-tech smartwatch that never sleeps. It's constantly listening to your body's secret language: the rhythm of your heart, the flow of your blood, the warmth of your skin, and the shake of your steps. This is the world of Wireless Body Area Networks (WBANs), a corner of science where tiny sensors talk to each other to keep an eye on your health. But here's the tricky part: your body is a chaotic, ever-changing place. A fast heartbeat is normal when you're running but dangerous when you're sleeping. A sudden drop in temperature might be a sign of fever or just a cold breeze. To spot a real health emergency, a computer can't just look at one single moment; it has to look at the story of what happened just before. This "story length" is called the temporal receptive field.

Think of the temporal receptive field like the zoom lens on a camera. If you zoom in too tight (a short field), you see the details of a single spike but miss the slow, creeping trend. If you zoom out too far (a long field), you see the big picture but miss the sudden, sharp danger. The big question scientists have been wrestling with is: How do you know which zoom level to use? Most existing systems are like a camera with a stuck lens; they pick one zoom level and hope it works for everything. But since your body changes its mood constantly, a stuck lens often misses the plot. This paper asks a bold question: What if the camera could decide for itself, in real-time, whether to zoom in or out, without needing a human to tell it how?

Enter ORCA, a clever new framework that acts like a smart, autonomous camera operator for your health data. The researchers propose that instead of forcing the computer to guess the right "zoom level" beforehand, we should give it a tiny, lightweight "supervisor" that watches the signals and instantly adjusts the focus. They call this agentic adaptation. It's not a giant, heavy brain that needs to be retrained every time your body changes; it's a nimble decision-maker that uses simple math tricks to say, "Hey, things are getting chaotic, let's look at the last 5 seconds!" or "Everything is calm, let's look back 1 second."

The team tested this idea on two very different stages. First, they used a custom dataset from wearable sensors, simulating the real-world chaos of a person moving around. Then, they threw ORCA into the deep end with MIMIC-IV, a massive, complex database of real hospital patient records. The results were impressive. ORCA didn't just guess; it learned to switch its focus dynamically. When the signals were stable, it used a short view. When an anomaly (a weird health event) started to brew, it automatically stretched its view to catch the longer pattern leading up to it.

Here is the magic: ORCA managed to perform just as well as the best "stuck lens" systems, hitting a score of 0.99 on a scale where 1.0 is perfect, but it did so without needing anyone to manually tune the settings first. In fact, on the hospital data, the adaptive version scored 0.9977 (AUROC), slightly beating the best fixed version which scored 0.9974. The paper suggests that this ability to change its mind on the fly is a game-changer for devices that need to run on tiny batteries and small chips, like the wearables we might wear tomorrow. It proves that you don't need a super-computer to be smart; you just need a system that knows when to look closer and when to step back.

The authors are careful to note that while ORCA is a strong contender, it isn't a magic wand that solves every medical mystery. It works best when the data is sliced into small, fixed chunks, and it relies on simple statistical rules rather than complex, learned policies. However, the findings strongly suggest that letting a model adapt its own "zoom lens" is a practical, robust way to handle the messy, non-stop drama of human physiology. By decoupling the heavy lifting of explanation from the fast-paced job of detection, ORCA offers a blueprint for future health monitors that are not only accurate but also light enough to live on your wrist.

Technical Summary: Agentic Anomaly Detection with ORCA

Problem Statement

Wireless Body Area Networks (WBANs) generate multivariate physiological time series that are highly nonstationary and must be processed under strict computational, memory, and energy constraints typical of edge devices. A critical, yet often overlooked, design challenge in this domain is the selection of the temporal receptive field—the amount of historical context used to determine if an observation is anomalous.

Physiological signals in WBANs exhibit rapid transitions between regimes (e.g., rest, activity, sleep, pathology). Consequently, anomalies are often contextual, meaning a signal deviation is only abnormal when interpreted relative to its surrounding temporal and cross-sensor context. Fixed-context models impose a static inductive bias that is rarely optimal across heterogeneous physiological regimes: short contexts may miss slowly evolving abnormalities, while long contexts dilute transient but clinically relevant deviations. Furthermore, heavy sequence models like Transformers often incur quadratic computational costs and high memory usage, making them unsuitable for resource-constrained wearable deployment.

Methodology: The ORCA Framework

The authors propose ORCA (Optimised Recurrent Convolutional Attention), a lightweight anomaly detection framework that introduces agentic control over temporal receptive fields. Rather than learning a single fixed inductive bias or using heavy attention mechanisms, ORCA dynamically adapts the effective temporal context at inference time based on lightweight signal statistics.

The architecture consists of three primary components:

Lightweight Supervisory Agent:
- This agent operates outside the gradient-based training loop and introduces no additional trainable parameters.
- At each time step, it computes inexpensive summary statistics (e.g., variance, autocorrelation) from the current input window.
- Based on these statistics, the agent autonomously selects a discrete temporal receptive field ( $k$ ) from a small set of permissible values. This selection is deterministic and based on the observation that stable regimes benefit from longer contexts, while abrupt shifts benefit from shorter ones.
Minimal Gated Recurrent Unit (MinGRU) Backbone:
- ORCA employs a simplified recurrent unit that eliminates the reset gate and recurrent affine transformations in the candidate computation to reduce overhead while preserving gated temporal memory.
- This backbone processes the multivariate sliding window to generate latent temporal representations.
Constant-Normalization Attention with Adaptive Receptive Field:
- Instead of standard Softmax, ORCA uses a constant-normalization operator (ConSmax), defined as $exp(S_{ij} - \beta) / \gamma$ . This removes sequence-level normalization dependencies, enabling fully parallelizable attention computation.
- A relative positional bias matrix is added to the attention scores, but its effective support is truncated to the receptive field $k$ selected by the supervisory agent. This restricts temporal interactions to a dynamically selected neighborhood without recomputing weights over the full window.

Post-Hoc Interpretability:
To support interpretability and short-horizon trend estimation without compromising the real-time detection pipeline, the framework integrates a Temporal Fusion Transformer (TFT) as a decoupled, post-hoc analysis module. This module is not part of the core edge-deployable detection loop.

Key Contributions

The paper outlines four main contributions:

Identification of Inductive Bias Limitations: The authors identify temporal receptive field selection as a critical inductive bias in WBAN anomaly detection and demonstrate the limitations of fixed-context modeling under nonstationary conditions.
Agentic Adaptation Framework: The introduction of ORCA, a lightweight framework that enables window-level adaptation of temporal receptive fields using inexpensive statistical probes rather than learned policies or additional parameters.
Empirical Validation: Validation of the framework's effectiveness and robustness on both a custom WBAN dataset and a large-scale clinical benchmark derived from MIMIC-IV.
Decoupled Interpretability: The provision of post-hoc interpretability and forecasting via a decoupled TFT module, ensuring the core detection pipeline remains edge-deployable.

Experimental Results

The authors evaluated ORCA against fixed-context baselines (with static $k \in \{1, 3, 5\}$ ) and other lightweight models.

Performance: On the custom WBAN dataset, ORCA achieved an AUROC of 0.9993, comparable to the strongest fixed-context baseline (0.9994) but without requiring manual tuning of the temporal horizon. On the MIMIC-IV benchmark, ORCA achieved an AUROC of 0.9983, slightly outperforming the best fixed-context configuration (0.9974).
Adaptation Behavior: Analysis of the supervisory agent revealed that it dynamically alternates between short, medium, and long receptive fields. Notably, during anomalous windows, the agent significantly favors longer receptive fields ( $k=5$ in 73.06% of anomalous windows vs. 39.88% in normal windows), aligning with the intuition that abnormal events often exhibit longer-range dependencies.
Efficiency: By replacing quadratic Softmax with constant-normalization attention and dynamically restricting the receptive field, ORCA reduces parameter count by over 6× compared to a full Transformer while maintaining linear attention complexity.
Ablation: Replacing the agent with a random policy or restricting it to single statistical cues (volatility or correlation alone) failed to produce systematic gains, confirming that the agentic control policy is essential for robustness.

Significance and Claims

The paper claims that adaptive temporal control is a practical and underexplored dimension for improving physiological anomaly detection in resource-constrained settings.

Practicality: ORCA demonstrates that high-performance anomaly detection does not require heavy sequence models or complex learned policies. By internalizing agentic decision-making within the temporal modeling process, the framework achieves robustness to nonstationarity without increasing model capacity or training instability.
Robustness: The framework maintains performance across heterogeneous physiological regimes and generalizes well to out-of-distribution clinical data (MIMIC-IV) without performance collapse.
Design Principle: The authors posit that temporal context selection should be treated as a dynamic, data-driven variable rather than a static hyperparameter. This approach bridges the gap between expressivity and efficiency, offering a robust design principle for edge-deployed anomaly detection in multimodal, nonstationary time series.

The paper remains modest regarding its limitations, acknowledging that the supervisory agent relies on deterministic heuristics which may not capture subtle regime transitions, and that the current formulation is limited to window-level binary classification rather than event-level localization or fine-grained severity estimation.