Sleep as a biological control circuit

Process S increases during wakefulness and decreases again during sleep. Process C describes the circadian influence on sleep and wake propensity. In simplified models, transitions occur when a state reaches time-dependent thresholds. Light can alter the circadian phase; sleep behaviour, in turn, influences when light is perceived. An original mathematical paper classifies these couplings as a hierarchical dynamical system. [1] The model explains temporal relationships, not every individual desire to sleep.

Pain, noise, heat, anxiety, and respiratory events can interrupt sleep, even though sleep pressure is present. Caffeine and medication alter further processes. For a technical bed, temperature or firmness are manipulated variables that only reach a part of the system. A measurement of movement or HRV is a noisy output signal and not a direct access to S or C. If every fluctuation is corrected, the bed itself may create disturbances through its interventions.

A useful controller requires limited manipulated variables, delay tolerance, and rules for uncertain signals. Stability here means, among other things, that small measurement errors do not trigger self-reinforcing counter-corrections. For biological models, it is additionally relevant whether they predict data outside the training sample. Models of performance under sleep loss show that short-term and cumulative effects are not always captured by the same simple relationship. [2] A nice curve progression is therefore no proof of validation. The accompanying page only represents a didactic model.

A system state is only reconstructible from measurements if the available signals contain sufficient information about it. However, several different causes can produce the same output signal: movement can indicate pain, awakening, or normal shifting of position. In that case, the cause is not uniquely identifiable from movement alone. Additional sensors only help if they actually provide independent information. For a smart bed, it should therefore be clarified in advance which disturbance can be detected at all. A complex control system with many parameters can otherwise only create pseudo-accuracy and react unstably in everyday use.

Filtering, feature extraction, and a classification model lie between the sensor and the visible sleep score. Every stage can lose information or introduce errors. A high correlation between two nocturnal mean values does not prove good agreement of individual events. Therefore, systematic deviation, scatter, and temporally matching comparison data are required for validation. The evaluation distinguishes between persons, nights, and short measurement epochs. Missing data must remain visible. A calm but awake person, poor skin contact, movement, or the signal of a partner can lead to plausible-looking misclassifications. For learning procedures, training and testing are separated by person. Otherwise, a model can recognise characteristics of the same person and overestimate its transferability. Devices and software versions are documented because an update can change the statement of an earlier validation.

Automatic adjustment requires more than a good sensor. It must be determined when a change is sensible, how large it may be, and when the system should suspend a decision. Filters and delays reduce noise but may obscure rapid changes. Frequent corrections can themselves generate noise, movement, or alertness. A good control loop therefore limits actuation speed and interventions and offers an understandable manual return to a stable state. For an internal trial, sensor errors, actuator movements, and sleep events are logged on the same timeline. The comparison includes an unmodified setup and an actively controlled setup with the most similar expectation possible. An increase in the automatically calculated sleep score is insufficient if the same algorithm drives the regulation and evaluates its success. Independent endpoints and a traceable data concept are necessary. Only the signals required for the purpose are stored; a product benefit does not presuppose the permanent disclosure of all raw data.