DATEXIS/sproto

S-Proto: Sparse Prototypical Networks for Long-Tail Clinical Diagnosis Prediction

Published at ECML PKDD 2024 (CORE A)
Boosting Long-Tail Data Classification with Sparse Prototypical Networks

Alexei Figueroa*, Jens-Michalis Papaioannou*, et al.
DATEXIS, Berliner Hochschule für Technik, Feinstein Institutes, TU Munich, Leibniz University Hannover
(* equal contribution)

This repository provides S-Proto, a sparse and interpretable prototypical network for extreme multi-label diagnosis prediction from clinical text. The model is designed to address the long-tail distribution of clinical diagnoses while preserving faithful, prototype-based explanations.

Interactive Demo

You can explore the model's predictions and interpretability features through our interactive web demo: https://s-proto.demo.datexis.com/

S-Proto was introduced in the paper:

Boosting Long-Tail Data Classification with Sparse Prototypical Networks European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD 2024, CORE A)
Alexei Figueroa*, Jens-Michalis Papaioannou*, et al.
DATEXIS, Berliner Hochschule für Technik, Feinstein Institutes, TU Munich, Leibniz University Hannover
(* equal contribution)

Overview

Clinical outcome prediction from Electronic Health Records is characterized by extreme label imbalance. A small number of diagnoses account for most patients, while the majority of diagnoses appear rarely. Standard transformer classifiers tend to perform well on frequent diagnoses but degrade sharply in the long tail.

S-Proto addresses this problem by extending prototypical networks with:

• Multiple prototypes per diagnosis
• Sparse winner-takes-all activation
• Prototype-level interpretability
• Efficient training despite increased representational capacity

The model achieves state-of-the-art performance on MIMIC-IV diagnosis prediction, with particularly strong gains in PR-AUC for rare diagnoses, and transfers successfully to unseen clinical datasets.

Model Architecture

S-Proto builds on PubMedBERT as the text encoder and introduces a sparse prototypical layer on top.

For each diagnosis label, the model learns multiple sub-networks, each consisting of:

• A label-specific attention vector
• A prototype vector representing a prototypical patient

Given an input clinical note:

1. The note is encoded using PubMedBERT
2. Token embeddings are projected into a latent space
3. Each diagnosis activates multiple candidate sub-networks
4. A winner-takes-all mechanism selects the single most relevant sub-network per diagnosis
5. Only the winning prototype contributes to the prediction and receives gradient updates

This allows S-Proto to model heterogeneous disease phenotypes while remaining sparse and efficient.

Intended Use

This model is intended for:

• Clinical diagnosis prediction from admission notes
• Research on long-tail learning in healthcare NLP
• Interpretable clinical decision support systems
• Analysis of disease phenotypes via learned prototypes

This model is not intended for direct clinical deployment without external validation, auditing, and regulatory approval.

Requirements

The model depends on the base sproto package (which contains MultiProtoModule) and specific versions of its dependencies. Version mismatches — especially in torchmetrics and pytorch-lightning — will cause AttributeError or import failures.

pip install torch>=1.12.1 \
            transformers==4.40.0 \
            torchmetrics==0.10.3 \
            pytorch-lightning==1.9 \
            huggingface-hub \
            matplotlib

Package	Required version	Reason
torch	>= 1.12.1	Minimum version for nn.PairwiseDistance and torch.einsum patterns used in the prototype layer
transformers	== 4.40.0	Required to bypass a metadata parsing bug
torchmetrics	== 0.10.3	MultilabelAveragePrecision was added in 0.10; older versions raise AttributeError on load
pytorch-lightning	== 1.9	MultiProtoModule is a pl.LightningModule; the exact API (e.g. validation_epoch_end) changed in 2.x
huggingface-hub	any	Required for fetching additional assets like thresholds and labels
matplotlib	any	Used for visualizations
sproto	bundled	The sproto/ package is included in this HF repo and downloaded automatically with trust_remote_code=True — no separate install needed

Inference Example

import torch
import sys
import json
from huggingface_hub import snapshot_download, hf_hub_download
from transformers import AutoTokenizer, AutoModel

def main():
    # 1. Download the repo and inject it into sys.path to resolve the internal 'sproto' package
    repo_id = "DATEXIS/sproto"
    repo_path = snapshot_download(repo_id)
    if repo_path not in sys.path:
        sys.path.insert(0, repo_path)

    # 2. Load Tokenizer and Model
    tokenizer = AutoTokenizer.from_pretrained(repo_id)
    # use_safetensors=False is required to bypass a metadata parsing bug in transformers 4.40.0
    model = AutoModel.from_pretrained(repo_id, trust_remote_code=True, use_safetensors=False)
    model.eval()

    # 3. Prepare Input Text
    text = """CHIEF COMPLAINT: depression, chest pain and vomiting
    
    PRESENT ILLNESS: The patient is a 53-year-old woman with a history of hypertension, diabetes, and depression. She developed severe anxiety and depression. She was having chest pains along with significant vomiting and diarrhea.
    """
    
    inputs = tokenizer(
        text, 
        return_tensors="pt", 
        padding="max_length", 
        truncation=True, 
        max_length=512
    )

    # Sproto requires raw token strings for its clinical section masking logic
    tokens = [tokenizer.convert_ids_to_tokens(ids) for ids in inputs["input_ids"]]

    # 4. Forward Pass
    with torch.no_grad():
        outputs = model(
            input_ids=inputs["input_ids"],
            attention_mask=inputs["attention_mask"],
            tokens=tokens
        )

    # Apply sigmoid to convert BCE loss logits to probabilities
    probs = torch.sigmoid(outputs.logits)[0]
    
    # 5. Fetch Labels and Thresholds dynamically from Hugging Face Hub
    try:
        labels_path = hf_hub_download(repo_id=repo_id, filename="labels.txt")
        icd_mapping_path = hf_hub_download(repo_id=repo_id, filename="icd_10_mappings.json")
        thresholds_path = hf_hub_download(repo_id=repo_id, filename="thresholds_per_label.json")
        
        with open(labels_path, "r") as f:
            labels = f.read().strip().split("\n")
        with open(icd_mapping_path, "r") as f:
            icd_mapping = json.load(f)
        with open(thresholds_path, "r") as f:
            threshold_mapping = json.load(f)
    except Exception as e:
        print(f"Warning: Could not load label mapping files from HF Hub: {e}")
        labels, threshold_mapping = None, None

    # 6. Evaluate and Print Results
    print("\n--- Inference Results ---")
    if labels and threshold_mapping:
        threshold_tensor = torch.zeros(len(labels))
        for idx, label in enumerate(labels):
            val = threshold_mapping.get(label, 0.20)
            threshold_tensor[idx] = val if val > 0.0 else 0.20 # Enforce valid > 0.0 threshold
        
        predicted_indices = torch.where(probs > threshold_tensor)[0]
    else:
        predicted_indices = torch.where(probs > 0.20)[0]

    if len(predicted_indices) == 0:
        print("No diagnoses predicted above the threshold.")
    else:
        results = []
        for idx in predicted_indices:
            idx_val = idx.item()
            prob = probs[idx_val].item()
            
            if labels and idx_val < len(labels):
                icd_code = labels[idx_val]
                description = icd_mapping.get(icd_code, "Unknown Description")
                results.append((icd_code, description, prob))
                
        # Sort alphabetically by ICD-10 code
        results.sort(key=lambda x: x[0])
        for icd_code, description, prob in results:
            print(f"- {icd_code} ({description}): {prob:.4f}")

if __name__ == "__main__":
    main()

Note: tokens (the list of token strings per sample) is required when use_attention=True (which is the default). The attention mechanism uses the actual token strings to mask clinical section headers ([CLS], [SEP], "chief complaint :", etc.) before computing token-to-prototype attention. Omitting tokens will raise a ValueError. Obtain them with tokenizer.convert_ids_to_tokens(input_ids[i]) as shown above.

Outputs

The model returns a dictionary with the following entries:

• logits
  Prediction scores per diagnosis label.

• max_indices
  Index of the winning prototype sub-network per diagnosis, corresponding to the selected prototype.

• metadata
  Additional information useful for analysis and interpretability.

Post-Processing & Label Thresholding

S-Proto outputs predictions over 1,643 individual ICD-10 diagnosis classes. To map the raw class indices back to actual diagnosis codes and filter out low-confidence predictions, you should use the thresholds_per_label.json file.

Because S-Proto was trained on highly imbalanced, long-tail data, using a single global probability threshold (e.g., > 0.50) will miss rare diseases. Instead, the model uses class-specific thresholds.

Important Thresholding Quirk: If a specific label has its threshold set to exactly 0.0 in the JSON file (which typically happens for extremely rare diseases with no validation positives), you should manually fall back to a reasonable default (e.g., 0.20) during inference. If you strictly apply probability > 0.0, the neural network's sigmoid function will falsely trigger for every patient.

Explainability

S-Proto provides built-in faithful explanations through its prototypical structure:

• Attention vectors highlight clinically relevant tokens
• Prototype distances reflect similarity to prototypical patients
• Multiple prototypes per diagnosis capture disease subtypes and cohorts
• Faithfulness metrics remain comparable to ProtoPatient despite higher capacity

Qualitative evaluation with medical professionals confirms that learned prototypes often correspond to clinically meaningful phenotypes.

Training

First, clone the repository:

git clone https://github.com/DATEXIS/sproto.git
cd sproto

Set up the environment using Poetry:

poetry install

Activate the virtual environment:

poetry env activate

Once the environment is active, you can start training by running the train command with the desired arguments.

Example:

train \
  --batch_size 3 \
  --pretrained_model microsoft/biomednlp-pubmedbert-base-uncased-abstract-fulltext \
  --pretrained_model_path path_to_pretrained_model.ckpt \
  --model_type MULTI_PROTO \
  --train_file training_data.csv \
  --val_file validation_data.csv \
  --test_file test_data.csv \
  --save_dir ../experiments/ \
  --gpus 1 \
  --check_val_every_n_epoch 2 \
  --num_warmup_steps 0 \
  --num_training_steps 50 \
  --max_length 512 \
  --lr_features 0.000005 \
  --lr_prototypes 0.001 \
  --lr_others 0.001 \
  --num_val_samples None \
  --use_attention True \
  --reduce_hidden_size 256 \
  --all_labels_path all_labels.pcl \
  --seed 42 \
  --label_column labels \
  --metric_opt auroc_macro \
  --train_files [] \
  --val_files [] \
  --only_test True \
  --model_name 5p \
  --store_metadata False \
  --num_prototypes_per_class 5

Citation

@inproceedings{figueroa2024sproto,
  title={Boosting Long-Tail Data Classification with Sparse Prototypical Networks},
  author={Figueroa, Alexei and Papaioannou, Jens-Michalis and Fallon, Conor and Bekiaridou, Alexandra and Bressem, Keno and Zanos, Stavros and Gers, Felix and Nejdl, Wolfgang and Löser, Alexander},
  booktitle={Proceedings of the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD)},
  year={2024}
}

License

This model and its associated code are released under the Apache License 2.0.

The model was trained on the MIMIC-IV dataset, which is subject to restricted access. No training data is included or redistributed with this repository. The data were accessed under a data use agreement. No patient-identifiable information is shared.

Use of this model must comply with all applicable data governance and ethical guidelines.

Limitations

• Extremely rare diagnoses remain challenging
• Clinical dataset biases may be reflected in predictions
• Winner-takes-all selection is fixed and not learned dynamically
• Not validated for real-world clinical deployment

Ethical Considerations

• The model processes sensitive clinical text
• Predictions should always be reviewed by qualified professionals
• Outputs should not be used as sole evidence for clinical decisions
• Care must be taken to avoid reinforcing existing healthcare biases