Context-Sensitive Modelling in Practice
A Study Design for Context-Sensitive Modelling in a Living Laboratory Setting
Overview
This post outlines a proposed study design for evaluating whether incorporating contextual structure into modelling improves predictive and explanatory performance in real-world healthcare settings. The study is situated within a voluntary and confidential, city-scale living laboratory that integrates clinical data, wearable measurements, and self-reported behavioural information.
The central question is whether models that incorporate both system state and contextual information provide more accurate and clinically useful inferences than models based on system state alone. This study proposal builds on a broader representational framework outlined in a previous post on this site.
Conceptual Framework
The study adopts a representational structure in which an individual at a given time is described by an extended state:
Ψ = (ψ, μ)
where:
ψ denotes system state, including clinical variables and physiological measurements
μ denotes contextual structure, including behavioural, environmental, and self-reported information
Observable outcomes O are modelled conditionally as:
P(O | ψ, μ)
The study evaluates whether inclusion of μ improves modelling performance relative to reduced representations based on ψ alone.
Study Design
Type of study
Prospective observational cohort study within a living laboratory environment.
Participants
Participants are recruited on a voluntary basis from the local population. Inclusion criteria are defined by the specific clinical domain under study, for example individuals with a chronic condition such as depression, chronic pain, or metabolic disease.
All participants provide informed consent and retain control over data sharing and system interaction.
Data Collection
Data are collected continuously and longitudinally from multiple sources:
System state (ψ):
clinical records where available
physiological measurements from wearable devices, such as heart rate, sleep, and activity
relevant biomarkers where applicable
Contextual structure (μ):
self-reported diary entries including mood, symptoms, and perceived stress
behavioural data such as activity patterns and routines
environmental context where available, for example time, location category, or social setting
treatment adherence and medication timing
Data are integrated via secure interfaces, with participants able to control when and how their data are accessed for analysis.
Modelling Approach
Two classes of models are constructed and compared:
Reduced models:
Models based on ψ only
P(O | ψ)Context-sensitive models:
Models based on extended states Ψ
P(O | ψ, μ)
Both model classes are trained on matched datasets and evaluated under equivalent conditions.
Outcomes
Primary outcomes depend on the clinical domain but may include:
prediction of symptom changes or flare events
classification of risk states
prediction of treatment response
functional or wellbeing measures
Performance is evaluated using standard metrics such as predictive accuracy, calibration, and error rates.
Evaluation Criteria
The principal evaluation criterion is whether:
P(O | ψ, μ) outperforms P(O | ψ)
This includes improvements in:
predictive accuracy
temporal precision of predictions
explanatory coherence with observed trajectories
clinical usefulness as assessed by practitioners
Illustrative Analytical Case
A key analytical focus is the identification of cases in which individuals exhibit similar system states ψ but different contextual structures μ, resulting in different outcomes O.
Formally, cases of interest satisfy:
π(Ψ₁) = π(Ψ₂) = ψ
but
P(O | Ψ₁) ≠ P(O | Ψ₂)
Such cases provide direct evidence that contextual structure contributes to outcome differentiation.
Ethical Considerations
The system operates on a voluntary and user-controlled basis. Participants determine when their data are used and for what purposes. The system is designed to support inference rather than prescribe behaviour, and to provide information that may assist both patients and clinicians in making more informed decisions.
Data governance, privacy protection, and transparency of use are integral to the study design.
Interpretation
The study does not seek to establish a comprehensive theory of human behaviour or health. Rather, it evaluates whether explicit representation of contextual structure improves modelling performance in practice compared to models that omit such context. The comparison is made under conditions in which contextual information is available through participant consent and engagement.
The study is therefore concerned with representational adequacy rather than ontological claims. Its aim is to determine whether models that incorporate contextual structure provide more accurate, coherent, or clinically useful inferences in real-world settings.
A positive result would support the view that omission of contextual structure leads to systematically incomplete models when outcomes depend on both system state and context. Conversely, the absence of improvement would suggest that contextual information, as represented in this setting, does not materially contribute to modelling performance.
In this sense, the study provides an empirical test of whether context-sensitive representation offers measurable advantages in applied modelling without presupposing a broader theory of human systems.
Conclusion
This study provides a concrete setting in which the role of contextual information in modelling can be evaluated empirically. By comparing reduced and context-sensitive models under controlled conditions, it aims to determine whether incorporating contextual structure leads to measurable improvements in inference and prediction to beneficial health effects.
The results may have implications not only for healthcare modelling, but for other domains in which outcomes depend on both system state and context.
