PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep learning models are not robust against noise in clinical text

Milad Moradi, Kathrin Blagec, Matthias Samwald

arXiv:2108.12242v1cs.CLcs.AI

TL;DR

Clean, human-curated benchmark datasets may not properly represent noisy real-world clinical inputs or reveal how robustly clinical NLP systems operate. The study introduces character-level and word-level perturbations and uses them to evaluate state-of-the-art NLP models across clinical tasks. Performance decreases across clinical NLP tasks when inputs contain noise.

  • Problem

    Clean, human-curated benchmark datasets may not properly represent noisy real-world clinical inputs or reveal how robustly clinical NLP systems operate.

    Source: §1 · p. 3

  • Method

    The study introduces character-level and word-level perturbations and uses them to evaluate state-of-the-art NLP models across clinical tasks.

    Source: §1 · pp. 3–4

  • Results

    Performance decreases across clinical NLP tasks when inputs contain noise.

    Source: §5 · pp. 8–9

  • Takeaways & Limitations

    Perturbation-based evaluation tests the robustness of clinical NLP systems using character-level and word-level perturbations.

    Source: §1 · p. 4, §7 · p. 13

    The paper does not evaluate how well the models perform on the clean benchmarks themselves, and many adversarial methods require access to model internals.

    Source: §1 · p. 4, §5 · p. 8

Abstract

from arXiv · show

Artificial Intelligence (AI) systems are attracting increasing interest in the medical domain due to their ability to learn complicated tasks that require human intelligence and expert knowledge. AI systems that utilize high-performance Natural Language Processing (NLP) models have achieved state-of-the-art results on a wide variety of clinical text processing benchmarks. They have even outperformed human accuracy on some tasks. However, performance evaluation of such AI systems have been limited to accuracy measures on curated and clean benchmark datasets that may not properly reflect how robustly these systems can operate in real-world situations. In order to address this challenge, we introduce and implement a wide variety of perturbation methods that simulate different types of noise and variability in clinical text data. While noisy samples produced by these perturbation methods can often be understood by humans, they may cause AI systems to make erroneous decisions. Conducting extensive experiments on several clinical text processing tasks, we evaluated the robustness of high-performance NLP models against various types of character-level and word-level noise. The results revealed that the NLP models performance degrades when the input contains small amounts of noise. This study is a significant step towards exposing vulnerabilities of AI models utilized in clinical text processing systems. The proposed perturbation methods can be used in performance evaluation tests to assess how robustly clinical NLP models can operate on noisy data, in real-world settings.

1. Introduction

Clinical NLP models achieve strong benchmark performance, but clean evaluations may not represent noisy real-world clinical text. This paper introduces perturbations and evaluates model robustness across clinical tasks.

  • Benchmark performance may not reflect how effectively clinical NLP systems operate on heterogeneous, noisy real-world inputs.Clinical text can contain misspellings, grammatical errors, omissions, duplication, abbreviations, acronyms, and jargon.Source: §1 · p. 3
  • The study evaluates robustness by applying character-level and word-level perturbations that simulate common clinical-text noise.The experiments cover four clinical NLP tasks and state-of-the-art NLP models.Source: §1 · pp. 3–4
  • Clinical NLP systems are sensitive to noisy data that humans can often understand, with performance decreasing when inputs contain noise.Source: §1 · p. 3
  • Unlike CheckList, the perturbation methods generate altered samples in the test set’s input space without persistent user intervention.They target both robustness and linguistic capabilities, whereas CheckList focuses on linguistic capabilities.Source: §1 · p. 4
  • The black-box perturbation methods can evaluate systems regardless of their underlying model, parameters, or target task.This avoids the internal-weight and model-structure access required by many adversarial attacks.Source: §1 · p. 4
  • The authors argue that benchmark evaluation should be complemented by tests revealing input regions where AI systems fail to operate accurately.Source: §1 · p. 4

2. Clinical NLP models

The study uses pretrained, domain-specific neural language models for clinical text processing. ClinicalBERT, ClinicalXLNet, and ClinicalELMo incorporate pretraining or further pretraining on biomedical or clinical text.

  • These models use unsupervised pretraining to encode language properties before supervised fine-tuning for downstream tasks.The described properties include lexical, syntactic, and semantic information.Source: §2 · p. 4
  • The clinical variants are adapted to biomedical or clinical text, including MIMIC-III clinical notes.ClinicalBERT and ClinicalXLNet are further pretrained on MIMIC-III notes, while ClinicalELMo is pretrained on MIMIC-III clinical texts.Source: §2 · pp. 4–5
  • ClinicalBERT, ClinicalXLNet, and ClinicalELMo are the three deep neural language models evaluated in the study.Source: §2 · p. 4

3. Clinical text processing tasks

The evaluation spans four clinical text processing tasks using datasets for semantic similarity, named entity recognition, relation extraction, and textual inference. Performance is assessed with task-appropriate metrics.

  • The study evaluates AI models on four clinical tasks: semantic similarity, named entity recognition, relation extraction, and textual inference.Source: §3 · p. 5
  • EBM PICO is a named entity recognition dataset containing almost 5,000 clinical trial abstracts annotated for populations, interventions, and comparators.Source: §3 · p. 5
  • i2b2-2010 is a relation extraction dataset with more than 400 patient reports annotated for eight relation types between medical problem and treatment entities.Source: §3 · p. 5
  • MedNLI contains sentence pairs from MIMIC-III clinical notes classified as entailment, contradiction, or neutral.Source: §3 · p. 5
  • ClinicalSTS-2019 contains more than 2,000 Mayo Clinic sentence pairs scored from 0 to 5 for semantic similarity.A score of 5 denotes semantic equivalence, while 0 denotes the opposite endpoint described in the passage.Source: §3 · p. 5
  • The evaluation uses accuracy, F1 variants, and Pearson correlation, with entity-level F1 for named entity recognition and correlation for semantic similarity.Source: §3 · pp. 5–6

4. Noisy inputs

The study constructs character-level and word-level perturbations to simulate noisy clinical text, while monitoring whether altered samples remain meaningful and label-consistent. The perturbations cover surface errors, lexical and grammatical changes, negation, and word-order variation.

4. Noisy inputs

  • Seven character-level and nine word-level perturbation methods were designed to simulate different types of noise in real-world clinical text.The methods were implemented as a broad evaluation set for noisy clinical inputs.Source: §4 · p. 6

4.1. Character-level noise

  • Character-level noise includes deletion, replacement, swapping, insertion, case changes, common misspellings, and character repetition.Deletion preserves the first and last characters, while insertion avoids the first and last positions.Source: §4.1 · p. 6
  • Additional character-level perturbations toggle letter case, replace words with common misspellings, or duplicate a selected character.Common misspellings are drawn from a Wikipedia corpus of frequently misspelled words.Source: §4.1 · p. 6

4.2. Word-level noise

  • Word-level noise replaces phrases with medical abbreviations, expands abbreviations, deletes or repeats words, substitutes synonyms, and changes number or tense.Abbreviations and acronyms come from a list of common medical and clinical forms, while synonyms are derived from frequent clinical words and distributional similarity.Source: §4.2 · pp. 6–7
  • Negation perturbations reverse positive and negative forms to test whether language models reflect negation in their decisions.The stated goal is to evaluate handling of negation rather than merely introduce lexical variation.Source: §4.2 · p. 7
  • Word-order perturbations change the order of M consecutive words to test sensitivity to word position versus the presence of important words.M is randomly selected from [2, |input|−1].Source: §4.2 · p. 7
  • Most perturbations preserve the original meaning and remain understandable, but deletion, negation, word order, and synonym replacement can make text meaningless or alter its meaning.Some altered samples may become inconsistent with the ground-truth label.Source: §4.2 · p. 7
  • The authors monitored the four meaning-changing perturbations, corrected labels when necessary, excluded unintelligible samples, and collected 200 samples per method and dataset.Character-level methods and word-level methods unlikely to change task-relevant meaning were used without filtering.Source: §4.2 · p. 7

5. Results and Discussion

Across four clinical text-processing tasks, neural language models lost performance on perturbed inputs, with word-level noise generally more damaging than character-level noise. The results also showed model- and task-specific robustness patterns, while examples demonstrated errors on understandable noisy text.

Model-specific effects

  • ClinicalXLNet generally achieved the highest results and was more robust to word-ordering noise, consistent with its permutation language-modeling strategy.The strategy allows context to be captured from important words regardless of their position.Source: §5 · pp. 8, 10

Overall results

  • Performance decreased on all clinical NLP tasks when inputs contained noise, as shown by evaluations on original and perturbed test sets.Table 3 reports model performance for character-level and word-level perturbations.Source: §5 · p. 9, Table 3 · p. 9

Noise sensitivity

  • Word-level perturbations caused larger performance losses than character-level perturbations across perturbation levels.Table 4 summarizes average absolute decreases for perturbation counts ranging from PPS=1 to 4.Source: §5 · p. 9, Table 4 · p. 10

Task-specific effects

  • Semantic similarity was especially affected because perturbed text changed the input representation and therefore the estimated similarity score.The reported mechanism concerns the representation used to calculate similarity from the altered text.Source: §5 · p. 9
  • Relation extraction was less affected than other tasks because decisions could rely on specific concepts remaining unperturbed.Important entity-related words had the greatest impact on the output in this task.Source: §5 · p. 9

Model-specific effects

  • ClinicalBERT handled synonym replacement more efficiently than the other systems, whereas ClinicalELMo performed worse overall.Figure 1 also shows ClinicalBERT changing from correct outputs on original samples to wrong decisions on noisy inputs.Source: §5 · pp. 8, 10, Figure 1 · p. 11

6. User study

The user study assessed whether perturbations remained understandable and preserved meaning, and whether they could be generated automatically. Most low-meaning-change perturbations preserved meaning, whereas meaning-changing perturbations often did not.

6. User study

  • Using Fleiss' kappa to assess inter-rater agreement, the study calculated a kappa value of 0.5054 from medically trained participants judging 50 noisy samples covering character-level and word-level perturbation methods.Source: §6 · p. 12
  • 87% of low-meaning-change perturbed samples were understandable and preserved the original meaning.These samples included all character-level methods and five word-level perturbations.Source: §6 · p. 12
  • 10% of low-meaning-change samples and 5% of meaning-changing samples produced tied judgments.Majority voting could not determine whether the perturbed text preserved meaning in these cases.Source: §6 · p. 12
  • Among the perturbed samples, 25% remained understandable but had changed meaning, while 40% were no longer understandable.There was a tie for 5% of the samples.Source: §6 · p. 12
  • The study supports automatic generation for many perturbations, but samples should be monitored, filtered, or corrected for meaningfulness and label consistency.Human supervision is especially needed when perturbations are expected to change meaning more frequently.Source: §6 · p. 12
  • Whether a perturbation changes meaning depends on the task, such as verb-tense changes affecting question answering but not necessarily disease-name NER.The authors therefore selected meaning-changing methods relative to the tasks studied, and this choice may differ for other tasks.Source: §6 · pp. 12–13

7. Conclusion

The paper introduces character-level and word-level perturbations to simulate noisy clinical text and evaluates high-performance clinical NLP systems. These models make erroneous decisions under slight noise, while the perturbations support robustness testing but vary in their need for manual validation.

7. Conclusion

  • The paper introduced character-level and word-level perturbation methods for simulating noisy inputs encountered in real-world clinical text processing.These methods are intended to support testing under noisy conditions.Source: §7 · p. 13
  • State-of-the-art clinical NLP models made erroneous decisions when inputs were slightly noisy.The models were evaluated on several clinical text processing tasks.Source: §7 · p. 13
  • The findings motivate future research on improving the robustness and reliability of AI systems intended for clinical decision support.The stated motivation concerns systems developed and tested for clinical decision support tools.Source: §7 · p. 13
  • Many perturbation methods can automatically generate noisy test samples, but some require manual validation for meaningfulness and consistency.The need for validation limits fully automatic use of those methods.Source: §7 · p. 13
  • The perturbations enable automatic and semi-automatic tests that reveal vulnerabilities in clinical NLP systems.The proposed tests are aimed at system developers, software testers, and end-users.Source: §7 · p. 13
Loading 2108.12242v1…