A Review on Human–AI Interaction in Machine Learning and Insights for Medical Applications

Mansoureh Maadi, Hadi Akbarzadeh Khorshidi, Uwe AickelinInternational Journal of Environmental Research and Public Health 2021 · DOI

TL;DR

This paper reviews the growing field of Human-in-the-Loop Machine Learning (HILML), particularly its application in medicine. It argues that combining human expertise with ML's computational power is crucial for developing reliable and trustworthy AI systems, especially in high-stakes domains like healthcare, and outlines key areas for future research.

What Problem Does This Paper Address?

As Machine Learning (ML) models become more sophisticated, they are increasingly applied to complex real-world problems. However, these models often struggle with issues like lack of transparency, the need for high-quality, unbiased data, and the inability to capture nuanced human understanding. This is particularly problematic in fields like medicine, where decisions have significant consequences. The paper addresses the challenge of effectively integrating human intelligence and domain expertise into ML workflows to overcome these limitations and build more robust, interpretable, and trustworthy AI systems.

Methods

This paper presents a scoping literature review. The authors searched Scopus and Google Scholar using keywords like "human-in-the-loop machine learning" and "interactive machine learning." They focused on peer-reviewed papers published between 2015 and 2020. To structure their review, they formulated four key research questions:

  1. "Why should humans be in the loop?"
  2. "Where does human–AI interaction occur in the ML processes?"
  3. "Who are the humans in the loop?"
  4. "How do humans interact with ML in HILML?"

The review categorizes human-AI interaction within the ML pipeline into three main stages: data producing and pre-processing, ML modeling, and ML evaluation and refinement.

Datasets

Not discussed in the paper. The paper is a review and does not present new experimental results or introduce new datasets.

Key Insights and Ideas

  • "Two heads are better than one": The core idea is that combining human domain expertise with ML's computational power leads to better outcomes than either can achieve alone.
  • ML's inherent limitations: ML models often fail when faced with complex, ill-posed problems, lack transparency (acting as "black boxes"), and require high-quality, unbiased data, which is not always available.
  • Human roles are diverse: Humans aren't just passive labelers; they contribute by providing data, selecting features, tuning parameters, interpreting results, and refining models.
  • Medical applications demand higher expertise: In high-risk fields like medicine, the expertise level of human collaborators is often higher, and their trust in ML systems is paramount.
  • HILML is a broad concept: Human-in-the-Loop Machine Learning (HILML) is presented as a more encompassing term than Interactive Machine Learning (IML), as it can involve human interaction at all stages of the ML process.

Key Results

The paper categorizes the reasons for human involvement in ML into three main areas:

  1. Tasks are too complicated: ML alone struggles with complex, real-world problems where human intuition and abstract thinking are needed.
  2. ML methods are not transparent and explicable: Humans can help interpret "black box" models, making them more trustworthy and understandable.
  3. ML results are not satisfactory for humans: Human subjective assessment and preferences are crucial for refining ML outputs to meet user needs.

The review also details where human-AI interaction occurs across the ML lifecycle:

  • Data Producing and Pre-processing: Humans provide data samples, label data, and clean data, often leveraging domain expertise.
  • ML Modelling: Humans contribute to feature selection, model creation (e.g., parameter tuning), and model selection.
  • ML Evaluation and Refinement: Humans evaluate ML outputs using subjective (satisfaction) and objective measures, and iteratively refine models to achieve desired results.

The paper also discusses the types of humans involved (domain experts vs. crowd workers) and the nature of their interaction (iterative, at specific times, or once).

Limitations and Future Work

The authors acknowledge that while HILML offers significant advantages, challenges remain. They highlight that the medical field, despite its critical need for reliable AI, has relatively few studies on HILML.

Promising future research directions suggested include:

  • Developing more sophisticated HILML approaches for medical data pre-processing, considering different data types and case studies.
  • Exploring interactive ML modeling techniques, such as rule-based ML and human-in-the-loop reinforcement learning, to better integrate medical expertise.
  • Designing user interfaces that facilitate effective human-AI interaction for ML evaluation and refinement, particularly for medical applications.
  • Further research into active learning for medical applications to reduce expert interaction costs and improve knowledge graph construction.
  • Investigating human-AI interaction for medical data pre-processing to improve data quality and leverage medical expertise.

Summary updated

Abstract

We design four questions to investigate and describe human-AI interaction in ML applications. These questions are "Why should humans be in the loop?", "Where does human-AI interaction occur in the ML processes?", "Who are the humans in the loop?", and "How do humans interact with ML in Human-In-the-Loop ML (HILML)?". To answer the first question, we describe three main reasons regarding the importance of human involvement in ML applications. To address the second question, human-AI interaction is investigated in three main algorithmic stages: 1. data producing and pre-processing; 2. ML modelling; and 3. ML evaluation and refinement. The importance of the expertise level of the humans in human-AI interaction is described to answer the third question. The number of human interactions in HILML is grouped into three categories to address the fourth question. We conclude the paper by offering a discussion on open opportunities for future research in HILML.
BibTeX
@article{maadi2021review,
  author = {Mansoureh Maadi and Hadi Akbarzadeh Khorshidi and Uwe Aickelin},
  title = {{A Review on Human–AI Interaction in Machine Learning and Insights for Medical Applications}},
  journal = {International Journal of Environmental Research and Public Health},
  year = {2021},
  doi = {10.3390/ijerph18042121},
}
Figure