Anxiety Symptom Trajectories Following AI-Powered Cognitive Behavioral Therapy in United Kingdom Primary Care: A Multilevel Growth Curve Analysis of the NHS Digital Wellbeing Programme

This prospective cohort study of 6,284 patients in UK primary care reveals that while AI-powered CBT significantly reduces anxiety symptoms on average, treatment response is heterogeneous across four distinct trajectory classes, with outcomes influenced by baseline severity, engagement, and a notable deprivation-related gap, ultimately supporting the platform's role as a scalable intervention in capacity-constrained areas.

The Gist

The Big Picture: A Digital "First Responder" for Anxiety

Imagine the UK's National Health Service (NHS) as a massive, busy hospital. For years, the waiting room for talking therapy (Cognitive Behavioral Therapy, or CBT) has been overflowing. People with anxiety are waiting months—sometimes over 90 days—to see a human therapist. It's like trying to get a table at a popular restaurant on a Friday night, but the waitlist is so long you might starve before you get seated.

To fix this, the NHS introduced a digital "first responder": an AI-powered app called CalmLogic. Think of this app as a 24/7 digital coach that teaches you how to manage anxiety using the same proven techniques a human therapist would, but instantly and for free.

This study asked three big questions:

1. Does this app actually work for people?
2. Does it work the same way for everyone?
3. Does where you live or how busy your local clinic is change how well it works?

The Experiment: A Race Against Time

The researchers looked at 6,284 people across 187 different GP practices in England. These people were anxious and signed up for the app. Over six months, they checked in regularly to report their anxiety levels (using a simple 0–21 score called the GAD-7).

The researchers didn't just look at the "average" result. Instead, they used a sophisticated statistical microscope (called Multilevel Growth Curve Modeling) to watch how each individual's anxiety changed over time, while also looking at the "neighborhood" (the GP practice) they belonged to.

The Results: Four Different "Race" Outcomes

If you just looked at the average, you'd see that anxiety went down for everyone. But the real story is that people ran this "race" in four very different ways. The researchers found four distinct groups:

1. The Sprinters (Rapid Responders - 28%):

   • The Analogy: These are the people who jumped on the app and saw huge improvements in the first two months. They were like runners who found a perfect tailwind.
   • Who they were: Often people with severe anxiety to start with, women, and those who really stuck with the app (completing most of the 10 modules).

2. The Marathoners (Gradual Improvers - 34%):

   • The Analogy: These people didn't sprint, but they kept a steady, consistent pace. They improved a little bit every week, like a slow but steady climb up a hill. By the end, they were much better off.
   • Who they were: The largest group. They had moderate anxiety and kept using the app regularly.

3. The Plateauers (Partial Responders - 23%):

   • The Analogy: These people started well, got a bit better, and then hit a wall. It's like a car running out of gas halfway up the hill. They improved, but not enough to feel "cured."
   • Who they were: They often had other issues like depression alongside their anxiety.

4. The Stalled Cars (Non-Responders - 15%):

   • The Analogy: For this group, the app didn't seem to help at all. Their anxiety stayed the same or even got slightly worse.
   • Who they were: This group had the highest rates of severe depression, long-term physical health problems, and were taking medication. They needed a different kind of help, not just an app.

The Surprising Twist: The "Busier" the Clinic, the Better the App Works

Here is the most interesting finding of the study.

Usually, you'd think that if a local clinic is super busy and has long wait times, the digital app would be a "last resort" and maybe work less well. But the study found the opposite.

• The Analogy: Imagine two towns. Town A has a short wait for a human doctor. Town B has a 6-month wait. In Town B, the people using the app improved faster than in Town A.
• Why? The researchers think that in Town B (where the wait is long), the app isn't just a "backup plan"; it's the only plan. People there are highly motivated to use the app because they have no other choice. In Town A, people might be less motivated because they know a human therapist is coming soon.

The takeaway: The app is most valuable exactly where the NHS is most broken. It acts as a crucial bridge for people who would otherwise be stuck in the waiting room for months.

The Inequality Gap: The "Deprivation" Hurdle

The study also found a sad but important truth about fairness. People living in the most deprived (poorest) areas improved more slowly than those in wealthier areas, even if they used the app just as much.

• The Analogy: Imagine two people running on treadmills. One is running on a smooth, flat track (wealthy area). The other is running on a track with sand and rocks (deprived area). Even if they both run at the same speed (same app usage), the person on the rocky track gets more tired and makes less progress.
• Why? The app can teach you how to manage anxiety, but it can't fix your housing instability, financial stress, or neighborhood safety. These real-life struggles make it harder for the "digital medicine" to work as well.

What Does This Mean for the Future?

1. It works, but it's not magic: The app helps the majority of people (over 60% got significantly better), but it won't fix everyone's anxiety alone.
2. We need to know who needs help faster: The study suggests we should look at who is using the app. If someone isn't improving after a few weeks (a "Plateauer"), the system should automatically flag them to see a human therapist sooner, rather than waiting months.
3. We need to fix the "rocky track": To make digital health fair, we need to pair these apps with extra support for people in poor communities, like community workers who can help with the real-life stressors that the app can't fix.

In short: The AI app is a powerful tool that is saving time and helping people, especially in areas where the NHS is overwhelmed. But to make it truly work for everyone, we need to recognize that some people need a little extra help to get over the hurdles of poverty and complex health issues.

Technical Summary

1. Problem Statement

The NHS Talking Therapies (formerly IAPT) programme faces severe capacity constraints, with median wait times for Cognitive Behavioral Therapy (CBT) exceeding 90 days in many areas. While AI-powered CBT platforms have been introduced as a scalable "Step 2" digital adjunct to bridge this gap, existing evidence is limited by:

• Lack of Longitudinal Data: Most studies rely on short-term RCTs with restrictive eligibility, failing to capture real-world trajectories over time.
• Homogeneity Assumption: Previous evaluations often report average treatment effects, obscuring the significant heterogeneity in individual patient responses observed in clinical practice.
• Contextual Blindness: There is limited understanding of how practice-level factors (e.g., local deprivation, service wait times) moderate the effectiveness of digital interventions.

2. Methodology

The study employed a prospective cohort design utilizing linked clinical, administrative, and platform data from the NHS Digital Wellbeing Programme.

• Sample: 6,284 patients (aged 18–65) from 187 general practices across four NHS England Integrated Care Systems (ICSs). Data was collected between April 2023 and September 2025.
• Intervention: CalmLogic, an MHRA-registered AI-CBT platform featuring a GPT-4 conversational agent, 10 structured CBT modules, and automated GAD-7 monitoring.
• Outcome Measure: Generalized Anxiety Disorder-7 (GAD-7) scores assessed at five time points: Baseline, Week 4, 8, 12, and 24.
• Statistical Approach:
  1. Three-Level Multilevel Growth Curve Modeling: To account for the nested structure (Assessments $\to$ Patients $\to$ Practices).
     • Level 1: Repeated measures (Time).
     • Level 2: Patient-level predictors (demographics, baseline severity, engagement, digital literacy).
     • Level 3: Practice-level predictors (deprivation index, list size, local IAPT wait times).
  2. Growth Mixture Modeling (GMM): To identify latent subpopulations with distinct trajectory patterns rather than assuming a single average trajectory.
  3. Predictor Analysis: Multinomial logistic regression (using a three-step approach) to examine predictors of class membership.
  4. Sensitivity Analysis: Multiple imputation (50 datasets) and comparison with two-level models to validate robustness.

3. Key Results

A. Average Trajectory and Variance

• Overall Decline: There was a significant average linear decline in GAD-7 scores of −0.94 points per month ($p < .001$), with a quadratic term indicating a decelerating rate of improvement over time.
• Heterogeneity: Substantial variance existed between patients in both intercepts (baseline severity) and slopes (rate of change).
• Contextual Effect: Practice-level factors accounted for 8.7% of the total variance in GAD-7 scores (ICC = 0.087), justifying the multilevel approach.

B. Latent Trajectory Classes (GMM)

Four distinct trajectory classes were identified:

1. Rapid Responders (28.4%): High baseline severity, steep decline stabilizing by Week 8. 71% achieved reliable recovery.
2. Gradual Improvers (34.1%): Moderate baseline severity with steady linear decline. 48% achieved reliable recovery.
3. Partial Responders (22.8%): Initial improvement followed by a plateau at clinically significant levels. Only 18% achieved recovery.
4. Non-Responders (14.7%): Minimal change or slight deterioration. High rates of comorbid depression and physical conditions. Only 4% achieved recovery.

C. Predictors of Response

• Patient-Level:
  • Positive Predictors: Higher baseline severity, female gender, higher digital literacy (eHEALS), and greater module completion predicted membership in the "Rapid Responder" class.
  • Negative Predictors: Comorbid depression and Asian/Asian British ethnicity (partially mediated by digital literacy) predicted slower trajectories.
  • Dose-Response: Each additional module completed was associated with an additional 0.11-point monthly decline in GAD-7.
• Practice-Level:
  • Wait Times: Counter-intuitively, practices with IAPT wait times >90 days predicted faster improvement trajectories ($\beta = -0.31$), suggesting AI-CBT provides the highest marginal benefit where traditional services are most constrained.
  • Deprivation Gap: Patients in the most deprived quintile showed significantly slower improvement trajectories ($\beta = 0.22$) despite equivalent engagement levels, indicating a deprivation-related treatment response gap.

4. Key Contributions

1. Methodological Advancement: Demonstrates the necessity of three-level modeling in digital health research, revealing that ignoring practice-level clustering (8.7% variance) risks inflated Type I errors.
2. Heterogeneity Mapping: Moves beyond "average effect" reporting to define a four-class taxonomy of response (Rapid, Gradual, Partial, Non), offering a nuanced framework for personalized care.
3. Contextual Validation: Provides empirical evidence that AI-CBT acts as a critical "bridge" intervention in capacity-constrained areas (wait times >90 days), validating its strategic placement in the NHS stepped-care model.
4. Equity Insight: Identifies a specific deprivation-related response gap, challenging the assumption that digital tools inherently democratize care and highlighting the need for targeted support in disadvantaged communities.

5. Significance and Implications

• Clinical Practice: The findings support a dynamic stepped-care approach. "Rapid" and "Gradual" improvers can remain at Step 2, while "Partial" and "Non-Responders" should be flagged early (e.g., via plateau detection algorithms) for step-up to therapist-supported care or specialist services.
• Policy & Commissioning: The data supports prioritizing AI-CBT deployment in Integrated Care Systems with the longest waiting lists, as these areas show the greatest incremental benefit.
• Equity Strategy: The identification of the deprivation gap necessitates the integration of non-digital support mechanisms (e.g., community navigators, peer support) to ensure digital interventions do not exacerbate health inequalities.
• Future Research: Highlights the need for pragmatic cluster-randomized trials and health economic evaluations (QALYs) to further validate cost-effectiveness and causal mechanisms.

Conclusion: AI-powered CBT in the NHS is effective on average but highly heterogeneous in individual outcomes. Its success is moderated by patient engagement, baseline severity, and crucially, the local service context. While it serves as a vital capacity bridge, specific strategies are required to address the slower response rates observed in the most deprived populations.