Highlights
– An AI-enabled decision aid (AI‑DA) that generated patient-specific digital twins improved decision quality (K‑DQI) and reduced decisional conflict and long-term regret compared with education alone.
– Patients randomized to the AI‑DA had substantially better knee‑specific function at 6–9 months (KOOS JR) and higher treatment concordance.
– No differences were observed in immediate shared decision‑making scores, 3‑month knee function, patient or clinician satisfaction, appointment duration, or TKA uptake.
Background
Knee osteoarthritis (OA) is a leading cause of chronic pain and disability worldwide and a common indication for total knee arthroplasty (TKA). Decision making about TKA is preference‑sensitive: the balance between potential benefits (improved pain and function) and risks (complications, recovery trajectory) varies across patients and depends on individual values. Shared decision making (SDM) and patient decision aids (DAs) improve knowledge, align treatments with patient values, and can reduce decisional conflict, but traditional DAs rarely incorporate personalized, dynamic prognostic information.
The concept of a “digital twin”—a computational, data‑driven virtual model of an individual that can simulate outcomes under different choices—promises to augment SDM by providing tailored risk:benefit predictions. The trial by Jayakumar et al. evaluated whether an AI‑enabled decision aid (AI‑DA) that created digital twins from PROMs and clinical data would improve decision quality, patient‑reported outcomes, and the SDM experience versus education alone in patients with knee OA considering TKA.
Study design
This was a randomized, open‑label clinical trial conducted at a single university‑affiliated orthopaedic clinic in the USA between February 2021 and November 2022 (ClinicalTrials.gov NCT04805554). Eligible patients had knee OA and were considering TKA. Participants were randomized to:
- Intervention: a full AI‑DA incorporating patient education, preference elicitation, and person‑specific benefit:risk predictions for TKA generated via digital twin models (n≈101 analyzed); or
- Control: patient education alone (n≈100 analyzed).
Primary outcome: Knee Osteoarthritis Decision Quality Instrument (K‑DQI). Secondary/process outcomes: CollaboRATE SDM measure, Decision Conflict Scale (DCS), Decision Regret Scale (DRS), KOOS JR (knee function) at 3 and 6 months, patient and clinician satisfaction, appointment duration, treatment concordance, and TKA rates. Follow‑up extended to 6–9 months for some outcomes.
Key findings
The trial randomized and analyzed 201 patients (intervention 101; control 100). Baseline characteristics were similar (mean age ~64 years; ~54–60% female).
Primary outcome: decision quality
Patients who used the AI‑DA reported higher decision quality on the K‑DQI (mean [SD] 84.4 [25.2]) than the education‑only group (71.4 [29.8]), P = 0.0011. This represents a statistically and clinically meaningful improvement in patients’ understanding of options, risks, and alignment of treatment with their goals.
Decisional conflict and regret
Decisional conflict (DCS) was lower in the intervention arm (1.0 [3.1]) than control (3.3 [5.8]); P = 0.0029. Decision regret at 6–9 months was also lower in the AI‑DA group (DRS 18.2 [19.5]) versus control (27.2 [24.2]); P = 0.0051. There was no statistically significant difference in regret at 3 months.
Knee function and clinical outcomes
Knee‑specific health measured by KOOS JR favored the AI‑DA at 6–9 months (69.5 [17.3]) versus control (47.0 [18.4]); P < 0.0001. The difference is sizeable and suggests improved medium‑term patient‑reported outcomes among those who used the personalized decision aid. At 3 months, KOOS JR scores were similar between groups.
Shared decision making, satisfaction, appointment length, and procedure rates
CollaboRATE scores (a brief measure of SDM) did not differ between groups, nor did patient or clinician satisfaction or appointment duration. Importantly, TKA rates were similar, indicating the AI‑DA changed the quality of decision making and subsequent functional outcomes without simply increasing or decreasing surgical utilization. Treatment concordance (the proportion of patients receiving care consistent with their stated preferences) was higher in the intervention group (91%) compared with control (76%), P = 0.0043.
Interpretation of effect sizes
The observed improvements in K‑DQI and reductions in decisional conflict and late regret indicate that individualized prognostic information enhanced patients’ decision processes. The large KOOS JR difference at 6–9 months suggests a potential downstream clinical benefit of higher‑quality, preference‑concordant choices; however, causality cannot be definitively established from a single center trial and the mechanisms merit further exploration.
Expert commentary and contextualization
Clinical context: SDM is a cornerstone of patient‑centered care for preference‑sensitive decisions such as TKA. High‑quality DAs consistently increase knowledge and decrease decisional conflict (Cochrane review). The present study advances the field by integrating AI‑driven, individualized prognostic modeling into the DA workflow—operationalizing the digital twin concept to provide person‑specific outcome trajectories during the clinical encounter.
How might digital twins add value?
Generic risk and benefit information does not account for patient heterogeneity. Digital twins can incorporate age, baseline function, comorbidities, prior PROMs, and other clinical variables to estimate likely outcomes and complications for that individual. In theory, more relevant predictions improve the accuracy of patients’ expectations and the alignment of care to their goals, both of which are linked to better patient‑reported outcomes and lower regret.
Strengths of the trial
- Randomized design with a pragmatic clinic‑based implementation.
- Use of validated measures of decision quality, conflict, and regret as well as patient‑reported knee function.
- Demonstrated improvements in both process (decision quality, concordance) and medium‑term outcome (KOOS JR) measures.
Limitations and generalizability
Key limitations include single‑center conduct and open‑label design. The precise AI models, their training data, and external validity are not fully described in summary form here; model performance across diverse populations requires independent validation. The unexpectedly large difference in KOOS JR at 6–9 months warrants cautious interpretation: potential explanations include (1) better matching of treatment to patient preferences, (2) differential adherence or engagement, or (3) unmeasured confounding. Replication in multicenter trials and assessment in more socioeconomically and racially diverse populations are needed.
Regulatory, ethical, and implementation considerations
Embedding AI prognostics in clinical encounters raises questions about transparency, interpretability, and responsibility for model errors. Effective implementation requires clear explanations to patients about model uncertainty, regular performance monitoring, and pathways for updates as new data accrue. Clinician training is also essential to integrate personalized predictions into shared deliberation rather than supplant clinician judgment.
Clinical implications and next steps
This trial suggests that AI‑augmented DAs using digital‑twin predictions can measurably improve decision quality and medium‑term function for patients considering TKA without increasing consultation time or altering surgical rates. For clinicians and health systems considering these tools, priorities should include:
- Independent external validation of predictive models across settings and populations.
- Evaluation of long‑term outcomes, cost‑effectiveness, and potential unintended effects (e.g., inequitable performance across subgroups).
- Development of standards for model transparency, patient communication of uncertainty, and integration with electronic health records.
Conclusion
The randomized trial by Jayakumar et al. demonstrates that an AI‑enabled decision aid generating patient‑specific digital twins can improve the quality of decisions and reduce decisional conflict and late regret among patients with knee OA considering TKA, with a notable improvement in 6–9 month knee function. These findings support further investigation of personalized, data‑driven decision support in preference‑sensitive surgical decisions, while underscoring the need for broader validation, transparent reporting of model performance, and careful implementation to ensure equitable benefit.
Funding and trial registration
Funding: Agency for Healthcare Research and Quality Grant (R21HS027037). ClinicalTrials.gov registration: NCT04805554.
References
1. Jayakumar P, Rathouz PJ, Lin E, Trutner Z, Uhler LM, Andrawis J, Koenig KM, Tsevat J, Bozic KJ. Shared decision making using digital twins in knee osteoarthritis care: a randomized clinical trial of an AI-enabled decision aid versus education alone on decision quality, physical function, and user experience. EClinicalMedicine. 2025 Oct 4;89:103545. doi: 10.1016/j.eclinm.2025.103545. PMID: 41112505; PMCID: PMC12528923.
2. Stacey D, Légaré F, Lewis K, Barry MJ, Bennett CL, Eden KB, Holmes‑Rovner M, Llewellyn‑Thomas H, Lyddiatt A, Thomson R, Trevena L. Decision aids for people facing health treatment or screening decisions. Cochrane Database Syst Rev. 2017 Apr 12;4(4):CD001431.
3. Elwyn G, Frosch D, Thomson R, Joseph‑Williams N, Lloyd A, Kinnersley P, Cording E, Tomson D, Dodd C, Rollnick S, Edwards A. Shared decision making: a model for clinical practice. J Gen Intern Med. 2012 Oct;27(10):1361–1367.
4. Barry MJ, Edgman‑Levitan S. Shared decision making — the pinnacle of patient‑centered care. N Engl J Med. 2012 Mar 1;366(9):780–781.
5. Topol EJ. High‑performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019 Jan;25(1):44–56.
AI image prompt for article thumbnail
A clinician and a middle‑aged patient seated at a clinic desk reviewing a tablet displaying a realistic 3D knee model and outcome probability graphs; the patient looks engaged, the clinician points to the tablet; warm, natural clinic lighting; photorealistic, high detail, professional medical setting.
