Randomized evidence on AI-enabled personalized treatment planning for adults with depressive–anxiety Spectrum disorders: A systematic review and Meta-analysis

Yıldız E., Yıldız E.

JOURNAL OF AFFECTIVE DISORDERS, cilt.1, sa.1, ss.1-13, 2026 (SCI-Expanded, SSCI, Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 1 Sayı: 1
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1016/j.jad.2026.121190
  • Dergi Adı: JOURNAL OF AFFECTIVE DISORDERS
  • Derginin Tarandığı İndeksler: Scopus, Science Citation Index Expanded (SCI-EXPANDED), Social Sciences Citation Index (SSCI), BIOSIS, CINAHL, EMBASE, MEDLINE, Psycinfo
  • Sayfa Sayıları: ss.1-13
  • İnönü Üniversitesi Adresli: Evet

Özet

Background

AI-enabled personalized treatment planning may improve outcomes by tailoring care, yet its clinical impact across modalities remains uncertain.

Methods

We preregistered a protocol (PROSPERO CRD420251106013), searched five databases, and included randomized trials comparing AI-enabled personalized planning (adaptive chatbots, clinician-facing CDSS/stratified care, individualized treatment rules [ITRs]) vs controls. Primary outcome was continuous symptom severity (Hedges' g; negative favors AI) at the longest follow-up using random-effects with Knapp–Hartung adjustment. Dichotomous response/remission was pooled as log OR.

Results

At prespecified longest follow-up, AI-enabled personalization reduced depressive symptoms (g = −0.35, 95% CI −0.63 to −0.08; k = 7; N = 757; I2 = 25.2%; τ2 = 0.0195; 95% PI −0.79 to 0.09) but not anxiety (g = −0.08, 95% CI −0.45 to 0.29; k = 7; N = 875; I2 = 46.2%; τ2 = 0.0548; 95% PI −0.76 to 0.61). Categorical outcomes favored AI (log OR = 0.35, 95% CI 0.01 to 0.69; OR ≈ 1.42; k = 4; N = 4398; I2 ≈ 0%). Effects tended to be larger vs passive comparators; evidence for clinician-facing CDSS and prospective ITR-guided allocation remains limited.

Conclusions

Patient-facing chatbots appear effective as adjuncts for depression but lack efficacy for anxiety, likely due to transdiagnostic limitations. Clinician-facing ITRs show promise for treatment matching but lack robust prospective validation. Notably, most trials pre-date Large Language Models (LLMs) (pre-2024); thus, conclusions regarding LLM-facilitated psychotherapy remain premature. Future trials should evaluate prospective model-guided allocation against active controls, include 6–12-month follow-up, and report variance statistics that enable synthesis.

Background

AI-enabled personalized treatment planning may improve outcomes by tailoring care, yet its clinical impact across modalities remains uncertain.

Methods

We preregistered a protocol (PROSPERO CRD420251106013), searched five databases, and included randomized trials comparing AI-enabled personalized planning (adaptive chatbots, clinician-facing CDSS/stratified care, individualized treatment rules [ITRs]) vs controls. Primary outcome was continuous symptom severity (Hedges' g; negative favors AI) at the longest follow-up using random-effects with Knapp–Hartung adjustment. Dichotomous response/remission was pooled as log OR.

Results

At prespecified longest follow-up, AI-enabled personalization reduced depressive symptoms (g = −0.35, 95% CI −0.63 to −0.08; k = 7; N = 757; I2 = 25.2%; τ2 = 0.0195; 95% PI −0.79 to 0.09) but not anxiety (g = −0.08, 95% CI −0.45 to 0.29; k = 7; N = 875; I2 = 46.2%; τ2 = 0.0548; 95% PI −0.76 to 0.61). Categorical outcomes favored AI (log OR = 0.35, 95% CI 0.01 to 0.69; OR ≈ 1.42; k = 4; N = 4398; I2 ≈ 0%). Effects tended to be larger vs passive comparators; evidence for clinician-facing CDSS and prospective ITR-guided allocation remains limited.

Conclusions

Patient-facing chatbots appear effective as adjuncts for depression but lack efficacy for anxiety, likely due to transdiagnostic limitations. Clinician-facing ITRs show promise for treatment matching but lack robust prospective validation. Notably, most trials pre-date Large Language Models (LLMs) (pre-2024); thus, conclusions regarding LLM-facilitated psychotherapy remain premature. Future trials should evaluate prospective model-guided allocation against active controls, include 6–12-month follow-up, and report variance statistics that enable synthesis.