🤖 machine learning

ProMUSE: Progressive Multi-modal Uncertainty-guided Staged Evidential Alzheimer Disease Classification

ProMUSE is a progressive, uncertainty-guided multi-modal network that adaptively incorporates expensive MRI or PET imaging only when necessary based on low-cost clinical data, achieving accurate Alzheimer's disease diagnosis while significantly reducing imaging costs and resource usage.

Original authors: Long Doan, Branden Chen, Ethan Litton, Huan Huang, Jiajing Huang, Yixin Xie, Weihua Zhou, Nandakumar Narayanan, Chen Zhao

Published 2026-06-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Long Doan, Branden Chen, Ethan Litton, Huan Huang, Jiajing Huang, Yixin Xie, Weihua Zhou, Nandakumar Narayanan, Chen Zhao

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

The Big Problem: The "All-or-Nothing" Medical Bill

Imagine you go to a doctor to check if you have a specific illness. Currently, the standard medical advice is: "To be 100% sure, we need to run every single test we have."

In the world of Alzheimer's diagnosis, these tests are:

Clinical Questions: A simple interview about your memory and health (Cheap and easy).
MRI Scan: A detailed picture of your brain's structure (Expensive).
PET Scan: A picture of how your brain is working and burning energy (Very expensive).

The problem is that most current computer programs designed to diagnose Alzheimer's assume you always have all three tests ready. They force the patient to pay for the MRI and PET scan every single time, even if the simple interview was enough to give a clear answer. This is like buying a full gourmet meal just to taste the appetizer.

The Solution: ProMUSE (The Smart Detective)

The authors propose a new system called ProMUSE. Think of ProMUSE not as a rigid machine, but as a smart, budget-conscious detective.

Instead of immediately calling in the expensive specialists (MRI and PET), the detective starts with the cheapest, most accessible tool: the Clinical Interview.

How the Detective Works (The Staged Approach)

ProMUSE uses a "progressive" strategy, meaning it adds tools only when absolutely necessary. Here is the step-by-step process:

Step 1: The Initial Clue (Clinical Data)
The detective looks at the patient's clinical data first. It asks: "Do I have enough clues to solve this case?"
- The "Uncertainty" Meter: The system has a special gauge that measures how "confused" it is. If the clues are clear, the gauge stays low. If the clues are vague, the gauge spikes high.
Step 2: The Decision Point
- If the gauge is low (Low Uncertainty): The detective says, "Case closed!" It makes a diagnosis using only the cheap clinical data. Result: The patient saves thousands of dollars because no scans were needed.
- If the gauge is high (High Uncertainty): The detective says, "I'm not sure yet. I need more help." It then brings in the next tool: the MRI Scan.
Step 3: The Second Opinion (MRI)
The detective combines the clinical clues with the MRI images. It checks the uncertainty gauge again.
- If it's clear now: Stop! Diagnosis made.
- If it's still fuzzy: The detective calls in the final, most expensive expert: the PET Scan.
Step 4: The Final Verdict
If the system reaches the PET scan, it combines all the evidence (Clinical + MRI + PET) to make the final call.

The Magic Glue: Dempster-Shafer Theory

You might wonder, "How does the detective combine a chat, a photo, and a metabolic scan without getting confused?"

The paper uses a mathematical concept called Dempster-Shafer Theory. Think of this as a "Confidence Blender."

Each test (Clinical, MRI, PET) gives the detective a "belief" (a level of confidence) and an "uncertainty" (how much it doesn't know).
The blender mixes these together. If the MRI agrees with the clinical interview, the confidence goes up, and the uncertainty goes down.
If the tests disagree, the system knows it's in a "conflict" zone and keeps the uncertainty high, prompting it to get more data.

The Results: Saving Money Without Losing Accuracy

The authors tested this "Smart Detective" on three major real-world datasets (ADNI, AIBL, and OASIS) involving thousands of patients.

The Savings: Because the system stops early when it's confident, it avoided using MRI and PET scans for a huge chunk of patients.
- In many cases, 50% to 90% of patients didn't need the expensive scans at all.
- The paper calculates that, on average, this saves a patient over $2,300 per diagnosis.
The Accuracy: Despite skipping the expensive tests for many people, ProMUSE was just as accurate as the systems that forced everyone to take all the tests.

Summary

ProMUSE is a new way to diagnose Alzheimer's that acts like a smart shopper. It starts with the cheapest option and only "buys" the expensive tests (MRI/PET) if the cheap option isn't enough to be sure.

Old Way: Buy everything, every time. (High cost, same accuracy).
ProMUSE Way: Buy what you need, when you need it. (Low cost, same accuracy).

The paper concludes that this method is a practical, money-saving solution for real-world clinics, though it notes that further testing on real human patients in a live hospital setting is still needed.

Technical Summary: ProMUSE

Problem Statement

Alzheimer's disease (AD) diagnosis increasingly relies on multi-modal data, including clinical assessments, structural Magnetic Resonance Imaging (MRI), and Positron Emission Tomography (PET). While combining these modalities improves diagnostic accuracy, the acquisition of MRI and PET scans is costly (ranging from ~$558 to $910 for MRI and ~$3,000 for PET) and not universally accessible. Existing prediction models typically assume full modality availability at inference time, implicitly requiring expensive imaging for every patient. This approach imposes substantial financial and clinical burdens, overlooking the practical constraints of real-world healthcare settings where cost and accessibility are critical factors. Furthermore, distinguishing Mild Cognitive Impairment (MCI) from Cognitively Normal (CN) individuals and full-blown AD is particularly challenging due to the subtle and heterogeneous characteristics of MCI.

Methodology

The authors propose ProMUSE (Progressive Multi-modal Uncertainty-guided Staged Evidential network), a framework designed to adaptively determine when additional modalities are necessary, thereby reducing data acquisition costs while maintaining diagnostic accuracy.

1. Data Representation and Pre-processing

Clinical Data: Processed using a Multi-Layer Perceptron (MLP) with four linear layers.
MRI and PET Data: Represented as graph-structured data ( $G = [E, X]$ $G = [E, X]$ ).
- MRI: Brain regions are segmented using FreeSurfer to extract radiomic features ( $X_{MRI}$ ), with edges ( $E$ ) defined by spatial proximity.
- PET: Uses the same node coordinates as MRI for anatomical consistency but derives node features ( $X_{PET}$ ) from the PET Unified Pipeline (PUP), including voxel counts and Standardized Uptake Value Ratios (SUVR).
- Processing: Both MRI and PET graphs are processed using a Graph Neural Network (GNN) architecture consisting of two GraphSAGE layers followed by a fully connected linear layer.

2. Uncertainty Quantification via Subjective Logic

ProMUSE employs a Dirichlet-based subjective logic model to quantify uncertainty for each modality independently:

Evidence Modeling: Output logits from each modality are transformed into non-negative evidence vectors using the Softplus activation function.
Dirichlet Parameterization: Evidence is mapped to Dirichlet parameters ( $d = e + 1$ ).
Belief and Uncertainty: The model calculates belief mass ( $B$ ) and uncertainty degree ( $U$ ) for each class. Uncertainty is defined as the unassigned probability mass, reflecting missing evidence. High-quality data yields higher evidence and lower uncertainty.

3. Multi-Modal Fusion via Dempster-Shafer Theory (DST)

To overcome the limitations of single-modality inputs, ProMUSE fuses evidence from multiple modalities using DST:

Fusion Rule: Belief masses and uncertainty degrees from different modalities (e.g., Clinical + MRI) are combined using the DST combination rule.
Conflict Handling: The fusion accounts for the conflict coefficient ( $C$ ) between modalities. High conflict increases uncertainty, while consistent evidence reduces it.
Normalization: The fused belief masses and uncertainty are normalized to ensure probabilistic consistency.

4. Staged Acquisition Strategy

The framework operates in a progressive, staged manner:

Stage 1 (Clinical): The model first analyzes low-cost clinical data. If the calculated uncertainty is below a learned threshold ( $\tau_1$ ), a final prediction is made.
Stage 2 (Clinical + MRI): If uncertainty exceeds $\tau_1$ , MRI data is acquired and fused with clinical data. If the new uncertainty is below a second threshold ( $\tau_2$ ), a prediction is made.
Stage 3 (Clinical + MRI + PET): If uncertainty still exceeds $\tau_2$ , PET data is added as the final modality to produce the prediction.

Threshold Selection: Optimal thresholds ( $\tau_1, \tau_2$ ) are determined via a data-driven grid search procedure to maximize accuracy on validation sets.

5. Loss Function

The model is trained using a hybrid loss function:

Cross-Entropy: For standard classification supervision.
Evidential Loss: Based on the Dirichlet distribution to encourage strong evidence for correct classes.
KL Divergence Regularization: Prevents the model from arbitrarily increasing evidence for all classes, with an annealing coefficient ( $I_r$ ) to control the regularization strength during training.

Key Contributions

Cost-Efficient Framework: ProMUSE introduces a staged acquisition strategy that prioritizes inexpensive clinical data, selectively incorporating expensive imaging (MRI/PET) only when uncertainty is high.
Uncertainty-Guided Decision Making: By leveraging Dirichlet-based subjective logic and Dempster-Shafer theory, the framework explicitly models and manages uncertainty to drive modality selection.
Resource Reduction: The approach significantly reduces reliance on high-cost resources without compromising diagnostic performance.
Comprehensive Evaluation: The framework is validated across three independent datasets (ADNI, AIBL, OASIS) and three binary classification tasks (CN vs. AD, CN vs. MCI, MCI vs. AD).

Experimental Results

Experiments were conducted on ADNI, AIBL, and OASIS datasets. ProMUSE was compared against 16 baseline models (including KNN, SVM, RF, and various deep learning architectures).

Accuracy: ProMUSE achieved competitive or superior accuracy compared to full-modality baselines.
- ADNI: 0.907 accuracy (CN vs. AD) and 0.810 (CN vs. MCI).
- AIBL: 0.945 accuracy (CN vs. AD) and 0.889 (CN vs. MCI).
- OASIS: 0.934 accuracy (CN vs. AD) and 0.990 (CN vs. MCI).
- Note: Performance on the MCI vs. AD task was generally lower across all models, including ProMUSE, reflecting the inherent difficulty of this classification.
Cost Savings: The staged strategy reduced MRI/PET usage by 50% to 90%.
- Cost analysis estimates an average saving of $2,325 to $3,558 per patient, depending on the dataset and task.
- For example, in the ADNI CN vs. AD task, 50% of patients were diagnosed using only clinical data, and 26.27% required only clinical + MRI, saving significant imaging costs.
Threshold Sensitivity: Analysis showed that lower uncertainty thresholds lead to higher modality usage. However, the authors note that lower uncertainty does not always guarantee higher accuracy, suggesting an optimal balance is required.

Significance and Claims

The paper claims that ProMUSE offers a practical, uncertainty-aware, and resource-efficient solution for real-world AD screening. By adaptively determining when additional modalities are necessary, the framework addresses the critical limitation of existing methods that assume full modality availability. The results demonstrate that high diagnostic accuracy can be maintained while drastically reducing the financial burden of data acquisition.

The authors explicitly state that while the results are promising, the framework has not yet been validated on real-world applications involving real human patients in a clinical workflow. They acknowledge that further testing is required to apply the framework in normal life and suggest that future work could focus on making the uncertainty degree more robust through additional algorithms.