Category: Blog

The hidden economics that are killing your device business

“Isn’t high device churn actually good for business? More churn means more sales, right?”

This question comes up in every executive meeting when discussing wearable retention strategies. On the surface, the logic seems sound: if users abandon their devices after 6 months, they’ll need to buy new ones, creating recurring hardware revenue. But this thinking reveals a fundamental misunderstanding of wearable economics that’s quietly destroying value across the industry.

The math tells a different story—one where high churn is actually a symptom of a broken business model, not a feature.

The Customer Acquisition Reality Check

The first crack in the “churn is good” theory appears when you examine customer acquisition costs. Recent data show that mobile app customer acquisition costs have surged 222% over the past decade, rising from $19 to $29 per user [1]. For fitness-focused applications, the numbers are even more sobering: acquiring an in-app buyer costs $74.68, while subscription customers cost $64.27 to acquire [2].
For fitness centers, the average customer acquisition cost sits at $118 per client [3]. When you extrapolate these figures to wearable devices—which require similar marketing investments, retail partnerships, and brand building—the true cost of replacement becomes clear.

Consider this: if you’re spending $50-150 to acquire each wearable customer (a conservative estimate based on industry benchmarks), and they abandon the device after 6 months, you’re in a constant battle to replace lost customers rather than building sustainable value from existing ones.

The Geographic Evidence

Recent market data reveals telling patterns about device adoption and abandonment across key markets:

India: The Low-Cost Trap India represents approximately 8.5% of population penetration (119 million units among ~1.4 billion people) but leads globally in total shipment volumes [4,5]. The market experienced 34% growth in 2023, reaching 134.2 million units [5]. However, industry analysis reveals that “watch lifecycles are as short as 6 months” due to “lackluster tracking performance leading to high abandonment” [6]. The smartwatch market actually declined 4.5% in 2024, partly due to an “influx of low-cost options” that created elevated inventory levels [7].

Singapore: The Engagement Success Story Singapore demonstrates the alternative model. With 44.1% fitness tracker ownership among older adults, driven by the National Steps Challenge providing free devices [8], the country shows how government partnership and engagement-focused strategies can drive sustained usage rather than churn.

UAE: The Premium Opportunity The UAE represents 30.2% of the Middle East & Africa wearable market, with 15% of the population using fitness trackers in 2023 [9]. The market shows healthy 14.3% CAGR growth, suggesting that focusing on engagement rather than replacement drives sustainable expansion.

Why the biggest opportunity in healthcare isn’t the next genomic breakthrough—it’s sitting in our pockets

Here’s an uncomfortable truth that should reshape every healthcare investment decision: we’re systematically investing in the wrong 20%.

While the healthcare industry pours $640+ billion annually into medical care and billions more into genomic research, mounting evidence reveals that medical interventions account for just 10-20% of health outcomes. Meanwhile, behavioral factors—which research consistently shows drive 40-50% of health outcomes—receive a fraction of the investment and attention.

This isn’t just an academic curiosity. It’s a massive market inefficiency that 2025 is finally beginning to correct.

The Evidence: What Actually Drives Health Outcomes

The data is overwhelming, even if the exact percentages vary by study. The seminal McGinnis & Foege research in JAMA identified behavioral factors as the leading “actual causes of death” in the United States. Subsequent analyses consistently confirm the hierarchy:

• 40-50% Behavioral factors (diet, exercise, substance use, medication adherence)
• 20% Social and environmental factors (income, education, housing, air quality)
• 20% Genetics (hereditary predispositions, family history)
• 10-20% Medical care (hospitals, drugs, procedures, devices)

Yet our healthcare spending is almost perfectly inverted. We dedicate massive resources to the 10-20% while largely ignoring the 40-50%.

The American Action Forum puts it bluntly: “95 percent of U.S. health expenditures go toward medical care,” while most experts agree that “medical services have a limited impact on health and well-being.”
This represents the largest ROI opportunity in healthcare—and 2025 is the year it’s finally being seized at scale.

In our previous post, we compared ID-Based and ID-Free recommendation models and found that ID-Free approaches generally produced higher-quality recommendations for personalized digital health nudging. In this post, we examine how ID-Free models behave under different conditions, focusing on three core questions:

1. How do ID-Free models scale as training data increases?
2. What trade-offs exist between recommendation quality, training time, and computational resources?
3. How much do semantic embeddings contribute, evaluated by comparing ID-Free and ID-Based versions of the same models?

To address these questions, we benchmarked several recommendation architectures powered by ID-Free semantic embeddings across varying training data sizes and conducted an ablation study contrasting each model’s ID-Free and ID-Based implementations.

Experiment Setup

We used the same proprietary digital health recommendation dataset as in our previous post. It consists of user–nudge interactions, enriched with metadata for both users (e.g., demographics, health conditions, aggregated tracker data) and nudges (e.g., content text, categories, target behaviors).

For all experiments, we maintained consistent validation and test sets, each covering one week of interaction data. The training data window was varied across 1, 2, 4, and 10 weeks to observe model behavior at different data volumes. Table 1 summarizes the dataset statistics for these splits.

The $185 billion wearables industry has a retention problem that threatens its entire value proposition. Here’s how AI-powered behavioral science can fix it.

The Engagement Cliff

Your Fitbit is probably in a drawer somewhere. You’re not alone—research shows that approximately 30% of wearable users abandon their devices within 6 months, with some studies documenting abandonment rates as high as 50% within just two weeks [1,2]. This isn’t a user problem; it’s a business model problem.

The current wearables paradigm treats engagement as a byproduct of hardware features. Companies invest billions in better sensors, longer battery life, and sleeker designs, then wonder why users lose interest once the novelty fades. Meanwhile, the real opportunity—transforming raw biometric data into sustained behavioral change—remains largely untapped.

The Data-to-Insight Chasm

Most wearables excel at data collection but fail spectacularly at insight generation. Your device knows you slept poorly, walked 3,000 steps, and had an elevated heart rate during your morning meeting. But it can’t tell you why these patterns emerged or what to do about them tomorrow.

This represents a fundamental misunderstanding of human motivation. While health outcomes are often simplified as 60% behavioral, 30% genetic, and 10% medical care [3], the key insight for wearables is that the vast majority of health improvement comes from modifiable behavioral factors—precisely the domain where current devices provide minimal value beyond basic tracking.
The gap isn’t technological; it’s methodological. Raw data doesn’t drive behavior change. Contextual narrative does.

The AI Transformation Evidence

2025 marks an inflection point where AI-powered behavioral interventions are moving from research labs to commercial deployment at population scale. The evidence is compelling:

Real-World Impact: CZ’s NudgeRank system, deployed across Singapore’s 1.1 million users, demonstrates that AI-driven personalized nudging achieves 6.17% increases in daily steps and 7.61% increases in exercise minutes—statistically significant improvements that persist across 12-week periods [4].

Industry Shift: Major players are pivoting toward AI coaching. Oura launched its AI Advisor, WHOOP deployed an AI coach, and Thrive AI Health (OpenAI’s collaboration with Thrive Global) is delivering hyper-personalized coaching at scale [5]. These aren’t experimental features—they’re core product strategies.

Economic Validation: Healthcare organizations implementing AI personalization are achieving 5-10% cost reductions while improving outcomes [6]. The business case extends beyond device sales to subscription revenue and B2B partnerships.

As behavioral health utilization reaches unprecedented levels, the intersection of AI-powered early intervention and reimbursement reform represents healthcare’s next critical evolution.

The Crisis Behind the Numbers

Healthcare is experiencing a behavioral health surge of historic proportions. Claims for inpatient behavioral health services jumped nearly 80% between January 2023 and December 2024, while outpatient services grew by 40% over the same period¹. This isn’t just a statistical anomaly—it represents a fundamental shift in healthcare demand that’s reshaping the industry’s cost structure and care delivery models.
The numbers tell a stark story of growing need, colliding with systemic barriers. Behavioral health spending has doubled over the past five years and now represents over 3% of total healthcare costs². One in three health plan actuaries now identify behavioral health services as a top cost inflator, with expected trends of 10-20% annually¹. Yet despite this surge in demand and spending, access barriers persist that push care toward the most expensive interventions.
The root of this paradox lies in a reimbursement system that hasn’t evolved to match the reality of behavioral health care delivery. Insurance reimbursements for behavioral health visits average 22% lower than for medical or surgical office visits³, creating systematic disincentives for provider participation and forcing patients toward crisis-level interventions that could have been prevented through earlier, less intensive care.

The next evolution of healthcare isn’t just about lowering copays – it’s about making those incentives psychologically compelling and actionable at scale.

The V-BID Boom Meets the Behavior Gap

Value-Based Insurance Design (V-BID) is expanding rapidly across healthcare. CMS has approved numerous Medicare Advantage V-BID demonstration programs that reduce or eliminate cost-sharing for high-value services like diabetes management medications, chronic disease monitoring, and preventive screenings. Employer-sponsored plans increasingly adopt similar models, recognizing that targeted financial incentives can theoretically improve health outcomes while managing long-term costs.

However, a significant gap exists between financial incentive availability and actual behavior change. Medication adherence rates for chronic conditions consistently remain around 50% across multiple systematic reviews, even when cost barriers are substantially reduced¹. Preventive care utilization often stays below optimal levels despite generous coverage improvements.

This disconnect reveals a fundamental challenge: financial incentives alone don’t reliably change behavior.

The issue isn’t the financial incentive design – it’s the missing behavioral layer that could transform economic motivation into sustained health actions.

Why Your Brain Doesn’t Care About Copay Reductions

Despite generous V-BID programs, medication adherence rates for chronic conditions remain around 50% according to systematic reviews¹. Preventive care utilization stays low even when cost-sharing is eliminated. The disconnect between financial incentives and actual behavior change reflects well-documented principles from behavioral economics research.

Thaler and Sunstein’s foundational work on choice architecture demonstrates that humans struggle with temporal discounting – the tendency to value immediate costs and benefits more heavily than future ones². A $30 copay reduction for diabetes medication might save $360 annually, but this abstract future benefit rarely competes with the immediate friction of complex medication schedules or side effects.

Research on health behavior interventions reveals several key factors for effectiveness:

Timing Matters: Just-in-Time Adaptive Interventions (JITAIs) show superior outcomes when delivered at optimal moments rather than on fixed schedules³. A medication reminder aligned with individual routines proves more effective than generic periodic messaging.

Personalization Drives Engagement: Systematic reviews of digital health interventions consistently find that personalized approaches achieve higher engagement rates than one-size-fits-all messaging⁴.

Loss Framing Effects: Meta-analyses confirm that loss-framed messages can be more motivating than gain-framed messages for certain health behaviors⁵.

Social Influence: Social comparison feedback has demonstrated effectiveness across multiple health behavior domains, though effects vary by population and context⁶.

Traditional V-BID programs reduce financial barriers but lack the behavioral intelligence to make those benefits psychologically compelling at the individual level.

In our previous post, “From IDs to Meaning: The Case for Semantic Embeddings in Recommendation,” we introduced the concept of moving beyond ID-based representations to metadata-driven approaches and why it’s promising for domains like digital health and personalized health, where personalization needs to go beyond static IDs. By leveraging semantic embeddings derived from descriptive metadata using Large Language Models (LLMs), ID-free approaches aim to address common challenges such as cold-start problems, limited generalizability, and deployment complexity.

In this post, we move from concept to evidence. Our central question is: How do standard recommendation architectures perform when powered by ID-free semantic embeddings, compared to our current ID-based Knowledge Graph Attention Network (KGAT) recommender model?

Models Compared

For this evaluation, we selected models that reflect two distinct approaches to user and nudge representation: Baseline Models (both non-embedding and ID-based) and ID-Free Enhanced Models.

Baseline Models

These models serve as foundational benchmarks for evaluating the value of more advanced, semantic embedding driven approaches:

• Random: A basic, non-personalized baseline that sets the lower bound for performance, whereby nudges are randomly selected for users.
• Popular: A simple heuristic that recommends items based on how frequently they were interacted with across the entire available history. This provides a static, global view of popularity and serves as a non-personalized baseline. While it doesn’t use embeddings or user-specific information, it offers a useful reference point for evaluating the performance of learned models.
• Knowledge Graph Attention Network (KGAT): Our current production model, KGAT, uses the traditional ID-based approach in a Graph Neural Network (GNN). It assigns unique numerical identifiers to each entity (e.g. users, nudges, markers, segments) within a knowledge graph and learns dedicated embeddings for each. An attention mechanism allows the model to capture complex, multi-hop relationships, making it especially effective at leveraging structured knowledge and explicit interaction histories, thereby serving as a strong ID-based benchmark [1]. This model has been successfully personalizing nudges for millions of participants daily, with very good results [2].

ID-Free Enhanced Models Leveraging Metadata

These models represent traditional, well-established recommendation architectures, but with a key change: instead of learning embeddings from arbitrary IDs, they are adapted to operate on pre-computed semantic embeddings. These embeddings are generated using LLMs applied to rich metadata—such as user health profiles, behavioral signals, and nudge content—allowing us to directly assess how well semantic inputs perform when integrated into widely adopted architectures. The following models were selected as they represented a spectrum of RecSys architectures from collaborative filtering to sequential recommenders.

• BPR (Bayesian Personalized Ranking): A classic pairwise ranking model for collaborative filtering, optimized for implicit feedback [3].
• NeuMF (Neural Matrix Factorization): A neural network-based approach for collaborative filtering, combining matrix factorization and multi-layer perceptron layers [4].
• SimpleX: A lightweight yet performant collaborative filtering model that incorporates historical user-item interaction sequences [5].
• SASRec (Self-Attentive Sequential Recommendation): A state-of-the-art sequential recommendation model that leverages self-attention over item sequences to capture dynamic behavioral patterns [6].

Experiment Setup

To ensure a robust and fair comparison, we conducted our experiments on our proprietary digital health recommendation dataset. The dataset encompasses 10 days of user interactions with various health nudges, alongside rich metadata for both users (e.g., demographics, health conditions, aggregated tracker data) and nudges (e.g., content text, categories, target behaviors, health nudges).

The following table summarizes the dataset statistics used for all models during training, validation, and testing.

Recommender systems have become indispensable for navigating the vast digital landscape, helping users sift through large volumes of content, products, and services. In the realm of digital health, they play an increasingly vital role in delivering personalized nudges and guidance, enhancing individual engagement, and supporting better health outcomes.

Traditionally, these systems rely on ID-based embeddings, where each user and item (or in our case, each individual and health nudge) is assigned a unique numerical identifier. These IDs are then converted into dense vector representations using techniques like matrix factorization or collaborative filtering. While effective for individuals with interaction histories and popular nudges, this approach has notable limitations, particularly around generalizability and cold-start scenarios.

Enter ID-free embeddings: a promising shift enabled by advances in Large Language Models (LLMs). Instead of relying on arbitrary IDs, these embeddings are generated directly from the rich, descriptive metadata associated with individuals and health nudges—think nudge content (e.g., text, language, tonality), individual health behavior data (e.g., fitness tracker data, dietary logs), demographic data, and health conditions. By leveraging the advanced semantic understanding capabilities of LLMs, ID-free embeddings capture the inherent meaning and relationships within this textual and behavioral information, offering a more flexible and robust foundation for personalized health recommendations.

Why Move Away from ID-Based Embeddings?

While conventional ID-based embeddings have been the mainstay of recommender systems for decades, it has several limitations:

1. Cold start: New users or nudges without historical data tend to perform poorly, as the system lacks enough context to generate meaningful recommendations. This leads to a poor initial experience for new users and slower adoption of new nudges.
2. Overfitting to history: Models often memorize frequent users or nudges, reducing their ability to generalize.
3. Limited portability: These models are transductive in nature, learning about specific users and nudges, and struggle to generalize to new or unseen ones. For example, user 123 in Deployment A of NudgeStream is a distinct individual from user 123 in Deployment B, so the models learned at each deployment cannot be easily transferred across deployments or use cases.
4. Operational complexity: ID-based pipelines often require custom retraining and tracking, since user and nudge IDs are unique to each deployment or population.

What is ID-Free Embeddings?

At their core, ID-free embeddings transform the descriptive text and structured data associated with individuals and health nudges into high-dimensional numerical vectors, using a pre-trained LLM encoder. For example, a nudge like “Need a snack? Choose crunchy veggies like carrots to curb hunger without the calories.” Is fed into the LLM, which then outputs a dense vector representing the semantic essence of that nudge. Similarly, an individual’s health profile, including their activity levels from a fitness tracker, dietary preferences, or health goals, can be encoded into an individual embedding. This process allows the recommender system to understand what a nudge is designed to achieve, or who an individual is, based on their attributes and behaviors, rather than just a unique identifier.

In CueZen’s recommender setup:

• User embeddings represent the individual’s current context, derived from a combination of static attributes (e.g. demographics) and dynamic attributes (e.g. recent health behaviors from tracker data).
• Item embeddings represent the nudge content itself. For example, a message like “a gentle reminder to take a 10-minute walk after lunch to boost energy” is encoded into a vector that captures its semantic meaning, such that nudges with similar messages or goals have similar embeddings. 

The promise of artificial intelligence (AI) in healthcare has always been tantalizing. We’ve seen AI make strides in diagnostics, drug discovery, and even in the management of chronic diseases. But one area where AI is poised to make an even more profound impact is in the realm of behavior change—specifically, how we can influence health behaviors at a population level. Enter NudgeRank™—a groundbreaking tool that harnesses the power of AI to deliver personalized health nudges on an unprecedented scale. Let’s dive into how this innovation works, why it matters, and what it could mean for the future of healthcare. 

The Herculean Task of Changing Health Behaviors

Changing patient behavior is one of the most formidable challenges in healthcare. Research shows that approximately 70% of premature deaths are linked to behaviors that can be modified, such as smoking, poor diet, and physical inactivity. For instance, smoking cessation alone could prevent 90% of lung cancers, but only about 7% of smokers succeed in quitting on their first attempt. Similarly, adherence to medication regimens is alarmingly low; studies estimate that nearly 50% of patients with chronic diseases in developed countries do not take their medications as prescribed, leading to increased hospitalizations and healthcare costs. These behaviors are difficult to change due to a complex interplay of factors—habits are deeply ingrained, motivation can diminish over time, and everyday life often presents barriers to maintaining healthy choices.

The World Health Organization (WHO) defines the burden of disease as the impact of a health problem measured by financial cost, mortality, morbidity, or other indicators. It reflects the gap between a population’s current health status and an ideal situation where everyone lives into old age, free of disease and disability. Behavior-related conditions like cardiovascular diseases, diabetes, and respiratory disorders contribute significantly to the global burden of disease. These conditions not only affect individual health but also place a substantial strain on healthcare systems, highlighting the critical need for effective behavior change interventions at a population level.

Traditional public health approaches—think mass media campaigns, pamphlets, and one-size-fits-all advice—often miss the mark. They fail to account for the unique circumstances of each individual, which is where the need for personalization becomes glaringly evident. What if we could tailor health advice to each person’s specific situation, habits, and needs? That’s the question NudgeRank™ is answering with a resounding “yes.”

NudgeRank™: Where AI Meets Personalized Health Interventions

NudgeRank™ is not your typical AI system. It’s a sophisticated digital nudging platform designed to influence health behaviors by delivering personalized, context-aware recommendations to millions of users. At its core, NudgeRank™ combines the predictive prowess of Graph Neural Networks (GNNs) with the dynamic adaptability of a Knowledge Graph, creating a system that learns, evolves, and optimizes its interventions over time.

Graph Neural Networks (GNNs): Graph Neural Networks are a powerful class of models that operate on graph-structured data, where entities (nodes) and their relationships (edges) can be represented in a more meaningful and flexible way than traditional grid-like data structures such as images or sequences. In NudgeRank™, the entities might be users, health interventions, behavioral goals, or nudges while the edges represent the relationships between these entities—how a user’s past behaviors might influence their future actions, or how certain interventions might be more effective based on demographic factors.

The essence of a GNN lies in its ability to perform message passing between nodes. Each node aggregates information from its neighbors, allowing the network to learn rich, context-sensitive representations of each entity based on higher-order relationships in the knowledge graph. In the context of NudgeRank™, this means that the system can learn which factors are most predictive of a successful health nudge. For example, a GNN might learn that users who have recently been more active but are starting to lapse respond well to motivational messages tailored to their previous high performance.

The GNN in NudgeRank™ uses a multi-layer architecture where each layer corresponds to a different level of abstraction. The initial layers might focus on simple features like the user’s age or activity level, while deeper layers capture more complex patterns such as the interaction effects between multiple health interventions. This hierarchical processing allows NudgeRank™ to build a nuanced understanding of each user’s unique circumstances.

The model’s attentive graph convolution layers are particularly crucial. These layers enable the GNN to weigh the importance of different nodes and edges, dynamically adjusting which information is most relevant in recommending a successful health nudge at any given time. For instance, if a user’s recent behavior suggests a decline in physical activity, the model can prioritize nudges that have historically been effective for similar patterns, ensuring that interventions are timely and contextually appropriate.

Knowledge Graph Integration: A Knowledge Graph is much more than a static database, it’s a dynamic structure that continuously evolves as new data is introduced. In NudgeRank™, the Knowledge Graph is populated with information from various sources—demographic data, historical health behaviors, interactions with past nudges, and even external data from wearables or health records. Each user’s data forms a personalized subgraph within the larger structure, with nodes representing their attributes and behaviors, and edges denoting relationships such as “increased activity after receiving a motivational nudge.”

What makes this integration so powerful is the extensibility of the Knowledge Graph. It’s designed to easily incorporate new types of data as they become available, allowing the system to adapt and improve over time. This is particularly important in healthcare, where new information—such as updates to a user’s medical history or the introduction of new health guidelines—can significantly alter the relevance of certain interventions.

Moreover, the Knowledge Graph is heterogeneous, meaning it includes a variety of node and edge types. For example, a node could represent a user’s demographic group, while another might represent a specific health goal like “increasing daily steps.” The edges between them could capture relationships such as how often the user met their step goals after receiving a certain type of nudge. This richness in representation allows NudgeRank™ to make more informed decisions about which nudges to prioritize, based on a deep, contextual understanding of the user’s situation.

The combination of GNNs and Knowledge Graphs in NudgeRank™ represents a significant advance in how we can personalize health interventions at scale. By leveraging the structure and relationships within the data, NudgeRank™ doesn’t just push generic health advice—it delivers finely-tuned, data-driven nudges that are optimized for each individual, adapting in real-time to changes in behavior and context.