Title : Clinical outcomes of AI-augmented cognitive rehabilitation in dementia and mild cognitive impairment
Abstract:
Background: Cognition-oriented treatments are associated with improved cognitive and non-cognitive outcomes among people with mild cognitive impairment (MCI) or dementia; however, access is limited by cost, availability of clinicians, and intervention frequency. Advances in artificial intelligence (AI) allow for scalable, frequent, and personalized cognitive interventions, while human-delivered telehealth interventions provide contextualized, goal-oriented, and psychotherapeutic support. This blended model may maximize clinical outcomes by combining the consistency and precision of AI with the expertise of human clinicians. We present data from a cohort of patients treated with this hybrid approach at the NewDays clinic, with a focused analysis of the MCI subgroup.
Methods: Thirty-seven patients (14 MCI, 19 dementia, 4 other cognitive symptoms; 17 males; mean age 78.8, SD = 10.0; mean education 16.5 years, SD = 2.4) were treated with the NewDays blended intervention approach and underwent evaluations at baseline and at 3–16 months (mean 7.4 months, SD = 3.5). Assessments included global cognition (ACE-III), verbal memory (WMS-IV Logical Memory I and II), meta-cognition (MMQ Satisfaction and Ability), and mood (GDS-15, GAI). For the MCI cohort, expected change without treatment was estimated from published trajectories, pro-rated to each patient's interval: 1.8 ACE-III points/year if stable and 5.3 if converting to dementia (Hamilton et al., 2024). To allow comparison with the larger MMSE literature, ACE-III scores were additionally converted to MMSE-equivalent scores (Matías-Guiu et al., 2018) and compared against an expected decline of 0.82 MMSE points/year (Xie et al., 2011). Patients were seen via telehealth by trained clinicians every two weeks, delivering individualized functionally-oriented treatment within a goal-focused reablement framework, while cognitive stimulation and training was delivered daily through conversational AI under clinician guidance. Results: Across all 37 patients, ACE-III Total scores improved from 72.4 to 74.8 (+2.4, p = 0.038), driven mainly by the memory subscale (+1.4, p = 0.021). Secondary measures improved significantly on Logical Memory I (+3.2, p = 0.005), Logical Memory II (+3.2, p = 0.005), MMQ Satisfaction (+2.6, p = 0.013), and GAI anxiety (−1.3, p = 0.022), while MMQ Ability (+1.5, p = 0.082) trended favorably. The MCI cohort (n = 14; 7 males; mean age 75.4, SD = 12.8; mean interval 8.3 months, SD = 3.3) on average did not decline further and in fact improved: observed ACE-III change was +3.1 points against an expected decline of −1.4 (p = 0.043), exceeding the no-treatment trajectory by 4.6 points, with 11 of 14 patients performing better than their expected trajectory. 6 of 14 exceeded their expected trajectory by 5 or more points. MMSE-equivalent scores showed the same pattern (observed +0.5 vs expected −0.6, p = 0.056). Logical Memory II improved by 4.8 points (p = 0.003; 13/14 improved or stable).
Conclusion: Our findings add further support for a blended human-AI treatment model in improving outcomes for people with cognitive impairment and extend it to MCI, where observed cognitive change exceeded the expected no-treatment trajectory. The single-arm design and modeled comparator limit causal inference. A randomized controlled trial (NCT07474038) is currently underway.
