Multi Day Fatigue Computation using Artificial Intelligence and a Single Sensor in an Uncontrolled Environment
Multi-day field events such as adventure races in the mountains and military missions require executive decision making, resilience, endurance and physical performance. Cognitive fatigue and physical fatigue are significant factors in decision-making, physical performance and safety. Fatigue has been studied for over a century with work performed to understand mechanistic results in the laboratory. The literature published on fatigue has concluded that fatigue is a complex multi-variate problem with systems including physiology, neurology and psychology where the context of the field is required to be fully representative. There are challenges moving into the field including logistics, validation, battery power, equipment size, assessment distractions from the mission task, validation protocols. Machine learning incorporates a process of collecting data and training models to reach a desired prediction accuracy. The model performance is increased by either tuning the model or adding new data to cover new scenarios the model has not previously observed. Wearable technology is capable of gathering appropriate data with minimal distractions compared to traditional laboratory assessments and machine learning is capable of analysing vast amounts of data. The thesis investigated if an AI model could accurately predict cognitive and physical fatigue using a new dataset from a single sensor with field induced noise from an unstructured environment including obstacles and terrain variation.
In the narrative literature review, the questions included: What is the definition of fatigue?, What sensor data are available from various positions on the human body?, and what cognitive and physical assessments can be used to validate field work? The key findings were that fatigue is a result of complex overlapping systems and the definition of fatigue requires context. A subset of available sensors was possible when considered against the limitations imposed by remote field environments and exercising participants. There were various cognitive and physical assessments possible that could be applied to field environments which covered different types of cognitive performance deficits. Having identified the definition, sensing and validation mechanisms for cognitive and physical fatigue the next step was to perform a field study to assess the practicality of the protocol.
In chapter 2 the first of two field experiments is described for a 12 day offshore ocean sailing voyage. The questions asked included: what cognitive assessments are sensitive to field loads, what tools and protocol is practical in remote multi day environments. A formula was derived to determine the contributing factors to compliance of an assessment protocol. Findings from this field study included; an assessment should be a single software tool to aid the researcher, compliance to protocols is challenging in a field environment with fatigued participants, it is possible to formulate compliance and use this as an approach when designing protocols, field protocols take time in logistics, data collection and analysis such that pilot studies are imperative to prove validity and logistics before scaling the protocol to multiple participants.
A set of assessment and cognitive load software tools were developed based on the learnings from the offshore sailing study. These were to stream line the participants experience and also aid the researcher with logistical challenges such as tracking data during the protocol.
A new protocol was designed for the second field study with the new software tools. The protocol allowed an hourly cycle of field load and assessments to take place. The aim of the second study was to determine if classification of human activity recognition (HAR) was possible in the field with a single sensor. A trail run was used as the new study with previous reported lab results as the comparison study. To the authors knowledge, this is the first study in the field with a single sensor and AI model using noise sources such as terrain, fatigue and self-pacing. Findings included a trail calibration protocol and data pipeline to process the raw sensor data to enable an AI model to be trained. The AI model was optimised to trade-offs between repetitive activities such as running compared to one off actions such as climbing a fence. Results matched previously published results from controlled environments with no terrain or fatigue variation (accuracy 97.8% vs 97.7% for trail vs lab).
The last experiment was to determine if fatigue could be modelled in a field environment. A second AI model was developed for regression and trained against the cognitive and neuromuscular assessments that showed greatest sensitivity to the protocol, that being finger tap test (FTT) and vertical jump. It was confirmed that the AI predictions using FTT had a mean absolute error (MAE) of 12.5% showing similar accuracy to laboratory based prior research.
In conclusion this thesis showed that fatigue is a complexity of redundant overlapping systems and a definition requires context from the field and mission goal to be applicable. Compliance is calculable and needs to be designed into the protocol to be successful in the field. A single sensor when coupled with a deep learning model can accurately classify human activity recognition. Fatigue can be predicted in the field using a single sensor and trained AI model. Further work is required to automate the data pipeline and protocol to enable multiple subject variation to be studied in the field.