ML-driven decision aid reduces decisional conflict in prostate cancer screening discussions compared to standard education.

This randomized controlled trial investigated the impact of a web-based, machine learning-driven decision aid on decision-making processes for prostate-specific antigen (PSA) screening. The study population consisted of middle-aged and older men aged 50 years and older, with a total sample size of 367 participants. The cohort was divided into two groups: the machine learning suggestion group (MLSG), comprising 185 participants, and the control group (CG), comprising 182 participants. The specific setting of the study was not reported in the available data. Participants in the MLSG received personalized PSA screening recommendations generated by the machine learning system. In contrast, the control group received standard video-based education, counseling, and values clarification without the generation of specific machine learning recommendations. The study design aimed to compare these two distinct approaches to shared decision-making support.

The primary outcome of the study was decisional conflict, measured using the total Decisional Conflict Scale (DCS) score. Results indicated that decisional conflict was significantly lower in the MLSG compared to the CG. The mean difference (MD) was -3.77, with a Cohen d effect size of -0.44. The 95% confidence interval for this difference ranged from -5.55 to -1.99, and the p-value was less than .001. Secondary outcomes included state anxiety, decision satisfaction, perceived support, adequate advice, decision confidence, worry, and calmness. Regarding perceived support (DCS7), the MLSG showed greater support with an adjusted p-value of .03. Adequate advice (DCS9) was also greater in the MLSG with an adjusted p-value less than .001. Decision confidence was higher in the MLSG for both DCS10 (adjusted P=.03) and DCS11 (adjusted P<.001).

Psychological outcomes were assessed using the State-Trait Anxiety Inventory (STAI). Total anxiety scores showed no difference between the two groups. However, specific items revealed distinct patterns. Worry, measured by STAI item 6, was reduced in the MLSG with a mean difference of -0.98 and a Cohen d of -0.89. The 95% confidence interval for worry reduction was -1.20 to -0.76, with an adjusted p-value less than .001. Calmness, measured by STAI item 1, was increased in the MLSG with a mean difference of 0.30 and a Cohen d of 0.25. The 95% confidence interval for this increase was 0.06 to 0.54, with an adjusted p-value of .01. Overall decision satisfaction, measured by the total Satisfaction with Decision score, was higher in the MLSG with a mean difference of -7.38. The 95% confidence interval was -8.54 to -6.18, and the p-value was less than .001.

The study also analyzed the alignment between machine learning suggestions and participant choices. When the machine learning suggested 'accept' screening, a significantly higher proportion of participants in the MLSG selected 'accept' compared to the CG (34 of 67 [50.7%] in MLSG versus 44 of 182 [24.2%] in CG), with a p-value less than .001. Conversely, when the machine learning suggested 'not now', the proportion of participants selecting 'accept' was lower in the MLSG than in the CG (21 of 118 [17.8%] in MLSG). The p-value for this comparison was not reported. Safety and tolerability findings were not reported in the study data, and no adverse events, serious adverse events, discontinuations, or specific tolerability metrics were documented.

Methodological limitations of this study include the lack of reported setting details and the absence of data on adverse events. A key limitation noted is that the psychological and behavioral effects of machine learning-assisted shared decision-making in geriatric populations remain poorly characterized. The study did not report funding sources or potential conflicts of interest. While the results support the scalable integration of AI-assisted decision support to foster patient-centered care in aging populations, the evidence is constrained by the incomplete reporting of safety data and the specific limitations of the machine learning model's performance in this demographic. The study does not establish long-term outcomes or the durability of these psychological benefits beyond the immediate decision-making process.

Clinical implications suggest that web-based, machine learning-driven decision aids may effectively reduce decisional conflict and enhance patient confidence regarding PSA screening discussions. The reduction in worry and increase in calmness observed in the machine learning group indicates potential benefits for patient anxiety management during screening decisions. However, clinicians should interpret these findings with caution given the lack of safety data and the limited characterization of behavioral effects in older adults. Further research is needed to address unanswered questions regarding long-term adherence, the generalizability of these findings to diverse geriatric populations, and the safety profile of such digital interventions. The current evidence supports the utility of these tools but highlights the need for more comprehensive safety monitoring and longitudinal studies.

Key takeaway: Consider ML decision aids to reduce conflict in PSA screening, though safety and long-term effects in older adults remain unclear.