Research

The Detection Problem

Modern medicine can intervene with increasing precision.

We can:

Yet intervention still follows visible outcomes.

The limitation is not intervention.

It is detection.

Physiological systems change before they fail.

But those changes are rarely observed directly.

What We Miss

Most health measurement is episodic.

These approaches detect:

They do not preserve:

What is lost is not data.

It is structure.

A Different Kind of Signal

Some physiological processes are not static variables.

They are continuous dynamics.

Respiration is one of them.

Each breathing cycle contains:

Across thousands of cycles per day, these patterns form a temporal signal.

Not a number.

A process.

Why Breathing

Respiration occupies a unique position in physiology.

It is:

This makes it both:

Across domains, respiratory patterns repeatedly appear as:

Examples include:

These observations are not unified.

But they are consistent.

What Makes It Observable Today

Until recently, continuous observation of respiration was impractical.

This has changed due to three converging factors:

Sensors — billions of smartphones with high-quality microphones capable of capturing airflow-related acoustic signals.

Computation — machine learning models capable of extracting structure from real-world audio.

Behavior — widespread acceptance of always-on sensing.

Respiration can now be observed using commodity hardware.

What Can Be Seen

From short recordings, it is already possible to extract:

Across recordings:

These observations are preliminary.

But they suggest that respiration may be treated as a structured signal.

What This Does NOT Mean

This does not imply:

Respiration is not a direct measurement of health.

It is a signal.

Its value depends on:

Many questions remain open:

These are areas of ongoing research.

References

The following references illustrate how respiratory dynamics repeatedly appear across clinical deterioration, neurological conditions, and longitudinal monitoring contexts.

Why Breathing Matters

Respiration is one of the few physiological processes that is simultaneously continuous, dynamically regulated, and coupled to multiple biological systems. It reflects interactions between autonomic regulation, metabolic demand, cardiovascular dynamics, neural control, and behavioral state. Unlike many physiological measurements that are captured intermittently through tests, devices, or clinical events, breathing unfolds continuously through time.

In most medical and consumer systems, respiration is reduced to isolated metrics such as respiratory rate. This compression removes much of the temporal structure contained within the signal itself. The framework argues that breathing should not be treated only as a vital sign, but as a structured physiological process whose organization may contain information beyond single measurements. The central proposition is not that respiration directly explains physiology, but that preserving respiratory structure may allow physiological change to become observable in a different way.

The framework distinguishes between physiological observation and physiological interpretation. Atum is framed not as a diagnostic system, but as an attempt to construct a continuous observation layer from respiratory structure. The goal is not to infer disease directly from breathing, but to determine whether respiration can function as a stable and computable representation of physiological state over time.

A recurring historical pattern appears throughout physiology: physiological systems often become more informative when continuous observation replaces snapshots. Electrocardiography transformed pulse into waveform structure. Continuous glucose monitoring transformed isolated glucose readings into trajectories, variability, and time-dependent patterns. Ambulatory blood pressure monitoring revealed nocturnal dynamics and hidden variability not visible in office measurements. The framework positions respiration within this broader historical pattern, while explicitly acknowledging that the respiratory case has not yet been validated to the same degree.

The argument therefore rests on a constrained hypothesis: if respiratory structure is both extractable and temporally stable, then breathing may function as a longitudinal physiological reference rather than a transient measurement.

Key findings

Limitations

References

Why Continuity Matters

A central premise of the framework is that continuity changes what becomes observable in physiology. Snapshot measurements capture isolated values and threshold crossings. Continuous observation preserves variability, oscillation, transitions, and temporal relationships across time. The distinction is treated as structural rather than quantitative: continuity is not defined as more frequent measurement, but as preservation of state through time.

The framework describes medicine as operating primarily through delayed indicators. Symptoms, biomarkers, imaging, and laboratory tests are generally interpreted as discrete events. Continuous physiological systems behave differently because they preserve trajectories rather than isolated points. ECG, CGM, and ambulatory blood pressure monitoring are used as examples where continuity exposed dynamics that could not be reconstructed retrospectively.

Within this framework, respiration is presented as an already continuous process whose structure is normally discarded. Respiratory rate preserves only a compressed summary of breathing activity. The proposed transformation is therefore not the creation of continuity itself, but the preservation of respiratory structure as a continuous signal.

If continuity is preserved, physiology may become observable as process rather than event. This distinction underlies the idea of longitudinal baselines. A baseline in this context is not a population average, but a reference derived from the temporal behavior of a specific individual. Deviation, drift, persistence, and trajectory all depend on continuity across time.

The concept of "time as moat" appears repeatedly, but always as a conditional outcome rather than an established property. The framework argues that longitudinal histories cannot be reconstructed retrospectively, and that model quality may depend on accumulated temporal depth. However, this remains dependent on unresolved questions surrounding temporal stability and repeatability.

Key findings

Limitations

References

Acoustic Extractability

The framework treats acoustic extractability as the first technical dependency in the system. Before any representation or longitudinal modeling becomes possible, respiration must first be recoverable from commodity hardware under realistic constraints.

Current implementation relies on smartphone microphones and short recording windows. The described system performs segmentation of inhale, exhale, and pause phases from audio recordings, followed by feature extraction and state representation. Internal results indicate that respiratory phases are recoverable with relatively high segmentation accuracy under controlled conditions. The framework treats this as evidence that breathing structure is computationally accessible.

The extraction pipeline is intentionally separated from downstream interpretation. The framework does not claim that successful segmentation proves physiological meaning. Instead, segmentation is framed as evidence that respiratory structure exists within the signal and can be computationally preserved.

Feature extraction currently produces multidimensional representations derived from temporal, phase, variability, structural, and acoustic characteristics. More than two hundred features are described across 10–30 second windows. The framework emphasizes that respiration appears not to be reducible to a single scalar metric.

Feasibility and validation are treated as separate thresholds. The current system has crossed a feasibility threshold, but not a validation threshold. The extraction layer exists and produces structured representations, yet robustness outside constrained environments remains unresolved. Real-world noise, device placement, uncontrolled environments, and longitudinal consistency are treated as open technical risks.

The framework therefore treats acoustic extractability not as proof of a physiological model, but as proof that respiratory structure can survive computational extraction.

Key findings

Limitations

References

Temporal Structure

Throughout the framework, respiration is treated as containing temporal organization beyond isolated measurements. This organization includes timing relationships, variability, transitions between phases, oscillatory behavior, and longer-range structure across recordings. The proposed "Breathing State" representation attempts to preserve this structure computationally rather than collapse it into summary metrics.

Current implementation represents respiratory recordings as vector embeddings over short windows. Internal observations suggest that recordings can exhibit clustering, repeatability, and partial separability across individuals and conditions. These observations are treated cautiously. The framework consistently distinguishes between early structure and validated state representation.

Temporal stability is presented as the central gating condition for the entire system. If respiratory structure cannot remain sufficiently coherent across time, then longitudinal baselines, trajectories, and accumulated histories lose meaning. The framework states that the system collapses into short-window feature extraction if persistence does not hold.

The proposed validation framework therefore focuses heavily on temporal behavior. Measurements include intra-person versus inter-person variance, clustering stability, embedding drift rate, and trajectory smoothness across contexts such as rest, activity, speech, and sleep. The central question is whether the same individual produces bounded and coherent respiratory structure across time while remaining separable from others.

Current status remains partial. The framework describes short-term structure and some constrained repeatability as supported, while long-term persistence, real-world robustness, and context invariance remain unresolved. Temporal structure is therefore treated as plausible and partially supported, but not validated.

Key findings

Limitations

References

Early Signal Dynamics

The framework proposes that respiration may respond earlier than many conventional physiological indicators because breathing is tightly linked to active regulatory systems rather than downstream outcomes alone. The claim is not that respiration uniquely predicts disease, but that respiratory dynamics may change before later-stage physiological failure becomes clinically visible.

This reasoning is partly grounded in broader physiological precedent. Continuous systems often reveal transitions and instability before discrete thresholds are crossed. The framework extends this logic to respiration by suggesting that breathing structure may preserve directional change, drift, or instability that snapshots fail to capture.

The proposed decision layer operates only on changes in respiratory structure over time. It does not attempt to infer biological meaning directly. Core computational primitives are intentionally limited to comparison, deviation, persistence, and trajectory. Outputs are framed as structural change relative to a personal baseline rather than diagnostic interpretation.

The framework maintains that early signal dynamics remain hypothetical until longitudinal validation exists. Current observations are limited to controlled recordings and short time horizons. No prospective clinical evidence demonstrates that respiratory continuity improves outcomes or enables validated intervention pipelines. The system therefore remains positioned as an observational infrastructure hypothesis rather than a proven predictive framework.

Key findings

Limitations

References

Physiological Coupling

The framework treats respiration as unusually connected to multiple physiological systems. Breathing is influenced simultaneously by autonomic regulation, metabolic demand, cardiovascular activity, neural control, sleep state, emotional regulation, and behavioral context. This multisystem coupling is presented as one reason respiration may preserve broader physiological information than isolated signals.

Importantly, the framework does not claim that respiration independently explains these systems. Instead, respiratory structure is framed as a convergence point through which multiple regulatory dynamics become partially observable. This distinction is repeatedly maintained throughout the material.

The proposed representation layer depends on the possibility that respiratory dynamics contain non-random and partially stable structure linked to these coupled systems. However, the framework explicitly acknowledges that non-redundancy relative to HRV and other physiological signals remains unproven. Whether respiration contributes uniquely informative structure beyond existing multimodal systems is still treated as a critical unresolved question.

The concept of personalized baselines emerges directly from this coupling logic. Because respiratory structure is expected to vary substantially across individuals and contexts, the framework rejects population averages as sufficient references. Baseline formation is therefore framed as longitudinal and individual-specific rather than normative.

Key findings

Limitations

References

Limitations

The framework consistently preserves uncertainty boundaries and explicitly identifies unresolved dependencies. This uncertainty structure is foundational to the architecture itself. Claims are repeatedly classified as proven, supported, observed, partially supported, or unproven.

Current limitations include small datasets, constrained environments, lack of longitudinal data, and incomplete cross-context validation. Existing evidence primarily demonstrates feasibility of extraction and existence of respiratory structure under controlled conditions. Stable longitudinal representation has not yet been demonstrated.

Several critical claims remain unresolved: long-term temporal stability, identity persistence, robustness to uncontrolled environments, context invariance, non-redundancy relative to other signals, meaningful longitudinal trajectories, stable baseline formation, and real-world passive monitoring feasibility.

The framework states that if temporal stability fails, the broader architecture collapses into short-window feature extraction without defensible longitudinal value.

Importantly, the system is not presented as a diagnostic or therapeutic platform. Interpretation layers remain external to the observation layer. The framework limits its scope to whether respiratory structure can become a stable computational substrate for observing physiological dynamics.

Key findings

Limitations

References

Research Roadmap

The proposed roadmap follows the dependency structure of the system itself. The framework emphasizes that higher-level claims are impossible unless lower-level conditions first hold. The sequence therefore progresses from extractability toward temporal persistence and only later toward longitudinal infrastructure behavior.

Current work focuses on robustness outside controlled environments, intra-person versus inter-person variance, baseline formation, identity versus state separation, and comparison against HRV and multimodal systems.

The broader research direction can be summarized as six sequential dependencies: (1) Extractability, (2) Structural stability, (3) State validity, (4) Non-redundancy, (5) Temporal advantage, (6) Dependency formation. These stages define the governing logic of validation.

Longitudinal validation is treated as the decisive threshold. The framework proposes that weeks-to-months datasets across multiple contexts will be necessary to evaluate clustering stability, drift behavior, baseline persistence, and trajectory formation. Only after these conditions hold could broader claims regarding reference layers, longitudinal accumulation, or infrastructure-like behavior become meaningful.

The roadmap therefore remains constrained and conditional. The system is presented not as a completed physiological model, but as an ongoing attempt to determine whether respiratory continuity can support a stable computational representation of physiological change over time.

References

The full research framework includes: validation architecture, claims hierarchy, temporal stability framework, representation logic, failure conditions, system boundaries, and supporting references.