Exploring Attitudes and Acceptance of Artificial Intelligence in Multiple Sclerosis from the Patient Perspective

Technical Summary: Exploring Attitudes and Acceptance of Artificial Intelligence in Multiple Sclerosis from the Patient Perspective

Problem Statement
Artificial intelligence (AI) is increasingly integrated into healthcare workflows, particularly for data-intensive chronic conditions like Multiple Sclerosis (MS) that rely on longitudinal monitoring and shared decision-making. While MS serves as a paradigm for data-driven care, the clinical translation of AI depends critically on patient acceptance and trust. Existing literature suggests that acceptance is heterogeneous, shaped by sociodemographic context, technology familiarity, and perceived risk. However, empirical data describing how people with MS (pwMS) perceive AI across distinct clinical use cases—ranging from administrative support to high-stakes diagnostic decisions—remains scarce. This study addresses the gap in understanding the specific attitudes, comfort levels, and predictors of acceptance among pwMS in Germany.

Methodology
The authors conducted a cross-sectional, web-based survey between December 2024 and April 2025.

• Participants: 241 adults (≥18 years) with self-reported MS of any subtype were recruited via digital channels (centre websites, social media, newsletters) and routine outpatient visits at two academic MS centers in Germany (Dresden and Düsseldorf).
• Instrument: A dedicated, self-administered questionnaire was developed, informed by the Technology Acceptance Model (TAM) and previous MS-specific instruments. It assessed four domains:
  1. Participant characteristics (demographics, disease course, self-reported disability via Patient-Determined Disease Steps [PDDS]).
  2. Technology access and usage frequency.
  3. AI literacy and exposure (self-rated knowledge, frequency of general and health-related AI use).
  4. Attitudes and preferences regarding AI in MS care across eight predefined clinical domains.
• Statistical Analysis:
  • An "AI attitudes composite" score was derived from preference outcomes, demonstrating high internal consistency (Cronbach's $\alpha$ = 0.90).
  • Associations were tested using Spearman's rank correlation, Mann-Whitney U tests, and Cochran's Q for differences across domains.
  • Multivariable linear regression was used to identify independent predictors of the composite acceptance score.
  • Multivariable proportional-odds (ordinal logistic) models were employed to analyze predictors for individual clinical domains.
  • Likert-scale outcomes were collapsed into three levels (negative, neutral, positive) to stabilize estimates.

Key Results

• Overall Acceptance: Acceptance was moderate overall (mean 3.39 ± 0.78 on a 5-point scale).
• Responsibility Gradient: Comfort levels varied significantly by clinical domain ($P < 0.001$). Participants were most comfortable with supportive, lower-risk applications: chronic management (54.4%) and symptom screening (50.2%). Acceptance dropped significantly for high-stakes decisions: treatment selection (38.6%) and diagnosis (35.3%).
• Decision-Making Preferences: When assuming equal diagnostic accuracy, 78.8% of participants preferred joint AI-clinician decision-making with the clinician retaining final responsibility. Only 0.8% preferred AI as the final decision-maker.
• Predictors of Acceptance:
  • Frequent General AI Use: The strongest independent predictor of acceptance (B=0.58, $P < 0.001$). Participants using general AI at least weekly showed significantly higher acceptance.
  • Region: Residence in Western Germany was associated with higher acceptance compared to Eastern Germany ($P = 0.025$).
  • Age: Older age was associated with lower acceptance of AI-supported chronic management.
  • Clinical Factors: Contrary to some hypotheses, clinical disability (PDDS) and disease duration were not significantly associated with acceptance in multivariable models.
• Demographics: Higher acceptance was observed in men, younger participants, and those with higher AI knowledge in unadjusted analyses, though only general AI use and region remained significant in adjusted models.

Significance and Claims
The paper posits that the successful integration of AI into MS care is primarily a sociotechnical challenge rather than a purely technical one. The authors claim that:

1. Acceptance is Context-Dependent: Patient comfort is not a binary "accept/reject" stance but follows a "responsibility gradient," where trust is higher for logistical/supportive tools and lower for autonomous clinical decision-making.
2. Familiarity Drives Adoption: Acceptance is more strongly linked to prior familiarity with AI (experiential exposure) than to disease severity or disability levels. This suggests that reluctance may stem from uncertainty about AI's governance and capabilities rather than a lack of interest or competence.
3. Implementation Strategy: The findings argue against a "one-size-fits-all" rollout. The authors advocate for a staged implementation strategy that begins with low-risk, high-acceptance use cases (e.g., monitoring, screening, administrative support) to build familiarity and trust.
4. Human-in-the-Loop Design: Given the strong preference for clinician final responsibility, AI tools should be designed as "human-in-the-loop" systems that augment rather than replace clinician judgment, with explicit accountability structures.

The study concludes that while technical validation is necessary, the priority must shift to designing transparent workflows and educational strategies that cultivate digital trust, particularly among older patients and in regions with lower baseline acceptance.