🧬 biology

Causal Responsibility Based Explainable AI for Vibrational Spectroscopy Applied to Cancer Diagnostics

This paper introduces Spec-ReX, a causal responsibility-based Explainable AI method for vibrational spectroscopy in cancer diagnostics, and rigorously validates its performance against other XAI techniques across in silico, in vitro, and ex vivo datasets to address the critical need for transparent clinical decision-making.

Original authors: Nathan Blake, David Kelly, Sarah Kapllani-Mucaj, Rong Wang, Geraint Thomas, Hana Chockler

Published 2026-06-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nathan Blake, David Kelly, Sarah Kapllani-Mucaj, Rong Wang, Geraint Thomas, Hana Chockler

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you have a super-smart robot doctor that can look at a tiny sample of human tissue and instantly tell if it's healthy or cancerous. It does this by analyzing "vibrational spectroscopy"—basically, listening to the unique "song" or "fingerprint" that molecules in the tissue sing when hit with light.

The problem? This robot is a black box. It gives you a diagnosis, but it won't tell you why. Did it hear a specific note? Did it ignore a certain sound? In medicine, knowing why is just as important as the answer itself, because doctors need to trust the machine.

This paper introduces a new tool called Spec-ReX to help us peek inside the black box. The authors compare Spec-ReX against two other popular tools (SHAP and Grad-CAM) to see which one does the best job of explaining the robot's reasoning.

Here is how they tested it, using three different "levels" of difficulty:

Level 1: The "Fake" Test (In Silico)

The Analogy: Imagine you are teaching a child to find a hidden treasure in a room. You secretly place a shiny red coin (the "ground truth") in a specific spot.

The Test: The researchers created fake data where they knew exactly which "notes" (wavenumbers) the robot was supposed to listen to.
The Result:
- Grad-CAM was like a child pointing at a whole corner of the room. It was right about the area, but too vague.
- SHAP was like a child pointing at the coin, but also pointing at the chair, the rug, and the window, saying, "It could be any of these!" It was sensitive but noisy.
- Spec-ReX was the most precise. It pointed only at the coin and ignored everything else. It was the "sniper" of the group, finding the exact cause with the least amount of distraction.

Level 2: The "Real Chemistry" Test (In Vitro)

The Analogy: Now, imagine the room is messy. The coin is still there, but it's buried under a pile of leaves, and the wind is blowing (this represents real-world noise and chemical changes).

The Test: They used real chemical mixtures in a lab. They knew DNA was present in some samples, but because of how the chemicals shifted and moved, the "DNA song" wasn't perfectly clear.
The Result:
- SHAP became very "sensitive." It found many places that might be DNA, casting a wide net.
- Spec-ReX became very "specific." It ignored the noise and focused only on the strongest, most reliable DNA signals.
- The Trade-off: The paper calls this a "sensitivity vs. specificity" tug-of-war. SHAP finds more potential clues (even if some are false alarms), while Spec-ReX finds fewer clues but is very sure about the ones it picks.

Level 3: The "Real Patient" Test (Ex Vivo)

The Analogy: Now we are in a real hospital. The "treasure" (the cancer signal) is hidden in a complex, living body. No one knows exactly where the treasure is, or even if it's a single coin or a whole pile of gold.

The Test: They looked at real tissue samples from the esophagus and mouth. There was no "answer key" to check against.
The Result:
- They found something surprising: Different robot brains think differently. Even when two robots were trained to do the exact same job, they looked at different parts of the tissue to make their decision.
- Spec-ReX showed that the robot was relying on very specific, sparse parts of the data.
- SHAP showed a broader, more scattered view.
- The Lesson: The paper warns that just because a tool highlights a specific note, it doesn't mean that note causes the cancer. It just means the robot thinks that note causes the diagnosis. The robot might be "cheating" (using a shortcut) rather than understanding the biology.

The Big Takeaway

The paper concludes that Spec-ReX is a fantastic tool for getting clear, focused, and sparse explanations. It tells you exactly which notes the robot is listening to without adding extra noise.

However, the authors are very careful to say:

No Magic Bullet: There isn't one "best" tool for every situation. Sometimes you want the wide net (SHAP), and sometimes you want the sniper scope (Spec-ReX).
Robot vs. Biology: These tools explain how the robot thinks, not necessarily how the disease works. A robot can be explained perfectly even if it's using a weird shortcut to get the right answer.
Human Check: Before doctors can trust these maps to diagnose patients, real human experts need to test them to see if they actually help in a real clinic.

In short, Spec-ReX is like a high-precision spotlight that cuts through the fog to show exactly what the AI is looking at, but we still need human doctors to decide if what the AI is looking at is actually the right thing to see.

Technical Summary: Causal Responsibility Based Explainable AI for Vibrational Spectroscopy Applied to Cancer Diagnostics

Problem Statement
The application of deep learning (DL) to vibrational spectroscopy (e.g., FTIR and Raman) for cancer diagnostics is increasing, yet these models often function as "black boxes." While they may achieve high classification accuracy, they fail to reveal which specific spectral features (wavenumber regions) drive their decisions. This lack of transparency hinders clinical trust and regulatory compliance (e.g., EU AI Act). Existing Explainable AI (XAI) methods like SHAP and Grad-CAM are often designed for generic use or specific data types (like images) and may not adequately address the unique requirements of spectral data, such as the need for local, feature-level explanations that distinguish between model behavior and biological causality.

Methodology
The authors introduce Spec-ReX, an XAI method adapted from the causal responsibility framework (ReX) to vibrational spectra. Spec-ReX quantifies the "causal responsibility" of specific spectral intervals for a model's classification using the formal theory of actual causality.

Mechanism: Spec-ReX employs a masking intervention strategy. It partitions a spectrum into contiguous intervals and replaces them with "neutral" values. Unlike standard perturbation methods that use minimum, mean, or linear interpolation, Spec-ReX utilizes a quadratic interpolation with adjustable Gaussian noise to join the ends of masked segments. This approach aims to minimize out-of-distribution (OOD) artifacts while removing spectral features.
Responsibility Calculation: The method calculates the degree of responsibility ( $r$ ) for an interval based on the size of the smallest "witness set" required to maintain the prediction. If an interval is sufficient on its own, its responsibility is 1; if multiple intervals are required, the responsibility is distributed ( $1/(k+1)$ ).
Evaluation Framework: The study evaluates Spec-ReX against Gradient SHAP and Grad-CAM across three distinct experimental stages:
1. In Silico (Definitive Ground Truth): Synthetic spectra with known, artificially added discriminative peaks.
2. In Silico (Ambiguous Ground Truth) & In Vitro: Spectra containing biochemical mixtures (specifically DNA vs. no DNA) where the model may learn a subset of relevant features, simulating real-world complexity and peak shifting.
3. Ex Vivo: Real tissue datasets (Oral FTIR and Oesophageal Raman) where no definitive attribution ground truth exists.

Key Contributions

Spec-ReX Implementation: The adaptation of causal responsibility theory to vibrational spectroscopy, providing sparse and specific attribution maps.
Systematic Benchmarking: A comprehensive comparison of Spec-ReX, SHAP, and Grad-CAM across synthetic, biochemical, and clinical datasets.
Metric Analysis: The application of diverse fidelity metrics, including Intersection over Union (IoU), Concentration, Positive Hit Rate (PHR), Insertion/Deletion AUC (iAUC/dAUC), and Fine-tuned Fidelity (FFid), to highlight the trade-offs between sensitivity and specificity in XAI evaluation.
Clarification of Causality: A rigorous distinction that XAI methods explain the trained model's behavior under specific perturbations, not biological causality or disease mechanisms.

Results

Definitive In Silico: Spec-ReX achieved the highest ground truth localization, with an IoU of 0.10 and an attribution concentration of 0.56, significantly outperforming the next best method (GradSHAP-datasetAvg with 0.26 concentration). Spec-ReX produced the sparsest maps, concentrating attribution mass almost exclusively on the defining peaks.
Ambiguous In Silico & In Vitro: A sensitivity–specificity trade-off emerged. SHAP demonstrated higher sensitivity, achieving the highest Positive Hit Rate (PHR) and identifying more DNA-associated regions. Spec-ReX was the most robust to the removal of low-importance regions (lowest FFid−), indicating it relies on a smaller, more specific set of features.
Ex Vivo (Clinical Data):
- Oesophageal Raman: Different model architectures (CNN vs. ResNet) learned different decision strategies, leading to different attribution patterns. Spec-ReX consistently produced sparser maps compared to the dense maps of SHAP and the broad regions of Grad-CAM.
- Oral FTIR: Applied to a PLS-DA model, Spec-ReX and SHAP performed similarly on insertion/deletion metrics, outperforming the Variable Influence on Projection (VIP) method. All methods agreed on the importance of the 1054–1062 cm⁻¹ region (carbohydrate/phosphate vibrations).
General Observation: No single metric or method was universally superior. Metrics rewarding sensitivity favored SHAP, while those rewarding specificity and robustness favored Spec-ReX.

Significance and Claims
The paper asserts that Spec-ReX provides sparse and specific model-causal responsibility maps for spectral classifiers. Its primary significance lies in its ability to identify the minimal set of features a model relies on, offering a more focused explanation than the often diffuse attributions of SHAP or the resolution-limited maps of Grad-CAM.

However, the authors maintain a modest stance on the implications:

Model-Centric, Not Biological: The explanations pertain strictly to the trained model's behavior under the masking intervention. They do not prove biological causality or that highlighted bands are causal disease mechanisms.
Utility of "Shortcut" Explanations: Even if a model relies on "shortcuts" (brittle or biochemically implausible features), Spec-ReX faithfully reveals this behavior, which is valuable for identifying model failure modes.
Clinical Validation Pending: The clinical utility of these sparse explanations remains to be validated through user-centered studies involving domain experts to determine if they improve trust calibration and diagnostic workflows.

The study concludes that while Spec-ReX is a competitive and effective method for explaining spectral classifiers, the choice of XAI tool should depend on the specific application and the trade-off between sensitivity and specificity required by the user.

Level 1: The "Fake" Test (In Silico)

Level 2: The "Real Chemistry" Test (In Vitro)

Level 3: The "Real Patient" Test (Ex Vivo)

The Big Takeaway

More like this