Guiding Neural Networks with Structured Domain Knowledge

198 concepts164 papers6 proposals

Deep learning models excel at finding statistical patterns in massive datasets. However, when applied to high-stakes, data-constrained domains like clinical medicine or fine-grained visual recognition, purely data-driven networks often stumble. They might overfit to spurious correlations, ignore subtle but critical diagnostic markers, or require unfeasibly large datasets to learn basic principles from scratch.

Human experts do not interpret data in a vacuum. A radiologist looking at a chest X-ray or a cardiologist reading an ECG relies on a rich framework of established rules, anatomical landmarks, and physiological mechanisms. The core idea of domain-informed knowledge integration is to inject this structured expert knowledge directly into neural network architectures. By constraining and guiding representation learning with human priors, we can build models that are more accurate, data-efficient, and inherently interpretable.

Why Purely Data-Driven Models Struggle

Consider an end-to-end convolutional network trained to classify retinal diseases. To the network, an image is merely a grid of pixels. It must discover entirely on its own that the spatial relationship between the optic disc and the macula is crucial, or that tiny red dots signify hemorrhages. In low-data regimes, the network might instead latch onto lighting conditions or hospital-specific camera artifacts.

As noted by Mehta et al., treating human specialists solely as providers of ground-truth labels discards the rich clinical concepts they use to differentiate complex pathologies. Similarly, Xiao et al. point out that purely data-driven models frequently fail to leverage structured domain expertise, limiting their ability to prioritize information according to established medical principles. When a model does not know what features should matter, it wastes capacity learning irrelevant noise.

Mechanism 1: Spatial and Structural Priors as Auxiliary Inputs

One of the most direct ways to guide a model is to provide expert-defined landmarks or regions of interest as extra input channels. Instead of forcing the network to locate critical structures from scratch, we explicitly hand them to the model.

For instance, Frank et al. demonstrated this by feeding explicit masks for the pleural line and vertical artifacts into standard neural networks for lung ultrasound analysis. By providing these domain-specific features as additional input channels alongside raw sensor data, off-the-shelf architectures could assess COVID-19 severity with significantly less training data.

In a similar vein, Goodfellow et al. incorporated domain-specific landmark likelihood maps—representing P, R, and T wave locations—as auxiliary input channels for ECG rhythm classification. This explicit spatial guidance drastically improves model convergence because the network doesn't have to deduce the coordinate system of cardiac cycles on its own.

Mechanism 2: Multi-Task Guidance and Standardized Criteria

Instead of modifying the inputs, we can shape the internal representations of a network by forcing it to predict intermediate, human-understandable concepts. This acts as a powerful regularizer, ensuring that the network's latent space aligns with expert reasoning.

Bône et al. applied this strategy to 3D CT scans for liver cancer diagnosis. They forced the model to first predict standardized clinical criteria (LI-RADS) before making a final classification. This architectural guidance yielded massive performance boosts over standard "black-box" models, regardless of the underlying backbone network.

This concept-guided approach is also highly effective in vision-language tasks. Mehta et al. used a large language model to generate a "concept bank" of medical features (like exudates or hemorrhages) to guide a vision-language model like CLIP, achieving superior few-shot performance. By translating established clinical diagnostic rules into measurable computational features—such as the "ABCD rule" for skin lesions explored by Yang et al.—models achieve better interpretability and performance in data-constrained environments.

Mechanism 3: Knowledge Graphs and Relational Attention

Domain knowledge often exists as structured relationships—such as drug-target interactions, anatomical hierarchies, or semantic attributes. We can embed these relationships directly into neural networks using graph neural networks or specialized attention mechanisms.

Wang et al. developed the Pharmacological Knowledge Graph Convolutional Network (PKGCN), which weights biological data based on known drug-target interactions and brain connectivity patterns. By integrating these established biological mechanisms into the network architecture, the model aligns its algorithmic focus with clinical expertise.

For fine-grained image recognition, Chen et al. introduced the Knowledge-Embedded Representation Learning (KERL) framework. KERL uses a structured knowledge graph of specific attributes (e.g., bird anatomical traits) to supervise low-level visual features, guiding the network to identify diagnostic details that standard convolutional networks miss. Similarly, Hong et al. designed a multilevel attention mechanism (MINA) that mimics how cardiologists read ECGs across different granularities (beat, rhythm, and frequency), making the model robust to real-world signal noise.

Mechanism 4: Hybrid Feature Fusion and Wide & Deep Architectures

Rather than relying entirely on deep neural networks to extract features, we can explicitly combine deep representations with handcrafted, expert-engineered features. This is often implemented via "Wide & Deep" architectures or contrastive alignment.

Natarajan won the 2020 PhysioNet/CinC challenge by pairing 22 handcrafted ECG features (such as specific wave morphologies and intervals) with a Transformer-based network. This hybrid approach leverages proven diagnostic indicators that are notoriously difficult for a network to learn from raw signals alone.

In another approach, Han et al. used contrastive learning to explicitly align deep image features from chest X-rays with quantitative, expert-driven radiomic features. This forced the neural network to prioritize the exact quantitative markers that human experts use for diagnosis. As Fox notes, incorporating domain-specific expert features serves as a powerful inductive bias that simplifies high-dimensional sensor data and significantly outperforms unsupervised representation learning in low-data environments.

Intuition: Why Domain Integration Works

Why does adding human rules improve a highly flexible neural network? The answer lies in the optimization landscape.

An unconstrained neural network has a massive hypothesis space. It can find millions of mathematical functions that perfectly separate the training data, but many of those functions rely on shortcuts or spurious correlations. By injecting domain knowledge—whether through auxiliary input channels, intermediate concept supervision, or graph-based attention weights—we alter the optimization landscape. We penalize solutions that violate established domain principles and reward solutions that align with expert logic.

Purely Data-Driven Search:
[Unconstrained Hypothesis Space] ---> Risk of overfitting to noise/spurious patterns

Domain-Guided Search:
[Unconstrained Space] + [Expert Priors/Rules] ---> [Constrained, Medically Valid Space]

This restriction of the search space effectively bridges the gap between raw data patterns and practical domain requirements. It allows models to achieve high diagnostic accuracy even with small datasets while providing human-readable justifications for their predictions.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_domain_informed_feature_engineering_e429eecc,
  title = {domain-informed feature engineering},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/domain-informed-feature-engineering/rdr_e429eeccbee345ff88fa94756d0ca569},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_domain_informed_feature_engineering_e429eecc_v1,
  title = {domain-informed feature engineering},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/domain-informed-feature-engineering/rdr_e429eeccbee345ff88fa94756d0ca569/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake