Receiving information on machine learning-based clinical decision support systems in psychiatric services increases staff trust in these systems: A randomized survey experiment

Erik Perfalk, Martin Bernstorff, Andreas Aalkjær Danielsen, Søren Dinesen ØstergaardmedRxiv 2024 · DOI

TL;DR

A simple educational intervention—providing basic information about how machine learning works—can significantly boost trust and reduce skepticism among psychiatric healthcare professionals. This randomized survey experiment demonstrates that even brief exposure to the mechanics of Clinical Decision Support Systems (CDSS) can help bridge the "implementation gap" in digital psychiatry.

What Problem Does This Paper Address?

Machine learning (ML) models often outperform clinicians in specific predictive tasks, yet they face a massive implementation hurdle in psychiatry: only about 0.4% of developed models actually make it into clinical practice.

A primary reason for this "valley of death" is a lack of trust from healthcare staff. If clinicians don't understand how a tool works or fear it might increase errors and inequality, they won't use it. While previous research showed that informing patients about ML could increase their trust, this paper asks: Can the same approach work for the doctors, nurses, and psychologists who actually have to use these tools?

Methods

The researchers conducted a randomized survey experiment involving 780 staff members from the Psychiatric Services of the Central Denmark Region. The study followed a rigorous experimental design:

  1. Baseline: Participants completed a questionnaire on their current work status, general trust in technology, and baseline knowledge of ML.
  2. Randomization: Participants were split into three groups:
  3. Intervention Arm: Received an information pamphlet (text and illustrations) explaining how ML-based CDSS works and how it aids clinical practice.
  4. Active Control Arm: Received a pamphlet describing the standard clinical decision-making process without ML.
  5. Blank Control Arm: Received no information.
  6. Outcome Measure: Participants answered questions regarding their trust and distrust in ML systems on an 11-point Likert scale.
  7. Analysis: The researchers used Principal Component Analysis (PCA) to group survey items into two distinct "sum scores": a Trust score and a Distrust score.

Datasets

Unlike a typical computer science paper, the "dataset" here consists of the survey responses from 780 healthcare professionals (out of 2,838 invitees). The cohort included a diverse mix of nurses (36%), psychologists (16%), social workers (17%), and doctors (14%).

The study design and analysis plan were pre-registered on the Open Science Framework (OSF), ensuring transparency and preventing "p-hacking."

Key Insights and Ideas

  • The "Explainability" Premium: The item regarding the importance of getting an explanation for a model's recommendation had the highest median score (10/10). This confirms that for clinicians, "black box" models are a non-starter; transparency is a prerequisite for trust.
  • Generalizability Across Stakeholders: The results almost perfectly mirror a previous study conducted on patients. This suggests that the psychological mechanisms of trust in AI are consistent across both the providers and recipients of care.
  • The Ceiling Effect: The intervention was more effective for older staff and those with bachelor-level education compared to younger or master-level staff. The authors hypothesize a "ceiling effect"—younger, more highly educated staff may already have a baseline level of trust or knowledge that leaves less room for improvement from a basic pamphlet.
  • Trust vs. Distrust: The study treated trust and distrust as two separate (though correlated) constructs rather than two ends of a single spectrum. The intervention successfully moved the needle on both: increasing positive trust while simultaneously lowering active concern.

Key Results

The intervention produced statistically significant improvements across the board:

  • Increased Trust: The intervention group showed a 5% higher trust score compared to both the active control (p < 0.001) and the blank control (p = 0.003).
  • Decreased Distrust: The intervention group showed a 3-4% decrease in distrust compared to the control groups (p = 0.004 to 0.021).
  • Perceived Understanding: Not surprisingly, the intervention group reported a significantly higher increase in their perceived understanding of ML systems (4.27% to 5.50% higher than controls).

While these percentages might seem small, in the context of a brief, 4-page pamphlet intervention, they represent a meaningful shift in professional attitude.

Limitations and Future Work

  • Selection Bias: The response rate was 27.5%. It is possible that staff who are already interested in (or skeptical of) technology were more likely to respond, potentially skewing the results.
  • Longevity: The study measured trust immediately after the intervention. It remains unknown if this "trust bump" persists over weeks or months of actual clinical practice.
  • Generic vs. Specific: The information provided was about ML in general. Future research should investigate if trust changes when clinicians are presented with a specific tool (e.g., a suicide risk predictor) versus a generic description.
  • Cultural Context: Denmark is one of the most digitalized countries in the world. The authors acknowledge that these results might differ in regions with lower baseline technological integration.

Summary updated

Abstract

Abstract Background Clinical decision support systems based on machine learning (ML) models are emerging within psychiatry. To ensure their successful implementation, healthcare staff needs to trust these systems. Here, we investigated if providing staff with basic information about ML-based clinical decision support systems enhances their trust in them. Methods We conducted a randomised survey experiment among staff in the Psychiatric Services of the Central Denmark Region. The participants were allocated to one of three arms, receiving different types of information: An intervention arm (receiving information on clinical decision-making supported by an ML model); an active control arm (receiving information on standard clinical decision process without ML support); and a blank control arm (no information). Subsequently, participants responded to various questions regarding their trust/distrust in ML-based clinical decision support systems. The effect of the intervention was assessed by pairwise comparisons between all randomization arms on sum scores of trust and distrust. Findings Among 2,838 invitees, 780 completed the survey experiment. The intervention enhanced trust and diminished distrust in ML-based clinical decision support systems compared with the active control arm (Trust: mean difference= 5% [95% confidence interval (CI): 2%; 9%], p-value &lt; 0.001; Distrust: mean difference=-4% [-7%; -1%], p-value = 0.042)) and the blank control arm (Trust: mean difference= 5% [2%; 11%], p-value = 0.003; Distrust: mean difference= -3% [-6%; - 1%], p-value = 0.021). Interpretation Providing information on ML-based clinical decision support systems in hospital psychiatry may increase healthcare staff trust in such systems.
BibTeX
@misc{perfalk2024receiving,
  author = {Erik Perfalk and Martin Bernstorff and Andreas Aalkjær Danielsen and Søren Dinesen Østergaard},
  title = {{Receiving information on machine learning-based clinical decision support systems in psychiatric services increases staff trust in these systems: A randomized survey experiment}},
  year = {2024},
  doi = {10.1101/2024.09.09.24313303},
  howpublished = {medRxiv},
}
Figure