Enhancing Animal and Human Healthcare Monitoring with Machine Learning: Overcoming Data Scarcity, Class Imbalance, and Lack of Generalisability

The proactive monitoring of health in both animals and humans, the early detection
of life-threatening diseases, and the ability to provide care in remote or
resource-limited settings are all hallmarks of a modern healthcare paradigm increasingly reliant on technological innovation. Central to this transformation is
the ever-evolving technology of wearable sensors and advanced computational
analysis. Specifically, a wealth of applications now leverage physiological and
behavioral time-series data, captured by ubiquitous sensors like accelerometers and
electrocardiograms (electrocardiograms (ECGs)), which are becoming foundational
building blocks of modern digital healthcare systems. The capacity to analyze the
vast and complex data streams provided by these sensors has been significantly
enhanced by the integration of artificial intelligence (AI). These intelligent systems
are not just tools for data processing; they are enabling a fundamental shift from
reactive to proactive and predictive health management. AI-driven innovations are
helping to manage the escalating complexity of health data, significantly improving
diagnostic accuracy and operational efficiency through data-driven analytics and
predictive insights. Furthermore, a key promise of this technological synergy lies
in the personalization of healthcare, paving the way for interventions and monitoring
strategies tailored to the unique biological and behavioral characteristics of
individual patients or animals.
This PhD thesis navigates this landscape through two primary and highly
relevant use cases: (i) animal health monitoring and (ii) human electrocardiogram
(ECG) classification. The first domain, animal health monitoring, is critical
for advancing animal welfare and enhancing the sustainability and productivity
of farming operations. To improve welfare, it is essential to first automatically
and accurately identify an animal’s behavior and then detect subtle deviations that
may signal illness, distress, or critical physiological events. This thesis explores
this challenge through activity recognition in goats and horses using accelerometer
data. Here, machine learning (ML) models are tasked with complex classification
problems, such as distinguishing between different gaits (e.g., walking, trotting,
cantering) in horses, classifying a range of behaviors in goats (e.g., grazing, fighting),
or detecting difficult to observe but critical events like the onset of parturition
in mares, where timely intervention can be life-saving. The second use case, human
electrocardiogram (ECG) classification, has garnered immense attention from the
xxiv SUMMARY
research community due to its potential to provide scalable, automated screening
for cardiac arrhythmias. The conventional method of manual interpretation by
cardiologists, while the gold standard, is time-consuming, requires specialized and
costly expertise, and can be a significant bottleneck in clinical workflows. ML
offers a powerful alternative, enabling the automated analysis of ECG signals to
detect abnormal heartbeats, such as supraventricular or ventricular ectopic beats.
This capability facilitates early and accessible screening, which is particularly vital
for improving healthcare outcomes in resource-limited regions where access to
cardiological expertise is scarce.
The application of ML in these real-world healthcare scenarios faces significant
and persistent challenges, which form the primary motivation for the research
contributions presented in this PhD. Specifically, this work confronts three critical
and often encountered issues that hinder the translation of ML models from
controlled research settings to robust on-farm or clinical deployment. The first is
(i) poor model generalization, a phenomenon where models trained on one group
of individuals fail to perform reliably on another. This is a direct consequence of
the high biological and behavioral variability inherent in living systems, stemming
from differences in genetics, age, physical build, and even sensor placement. The
second major obstacle is (ii) the scarcity and high cost of acquiring large, accurately
labeled datasets. Supervised ML models typically require vast amounts of annotated
data to learn effectively, yet the process of collecting and labeling health data
is exceptionally resource-intensive, often requiring expert human annotators (e.g.,
veterinarians or cardiologists) and navigating strict privacy and ethical regulations.
The third critical issue is (iii) the difficulty of learning from imbalanced data. In
health monitoring, the events of greatest interest, such as an arrhythmia or the onset
of a disease, are often rare compared to the vast amount of data corresponding
to a normal state. Standard algorithms, optimizing for overall accuracy, tend to
develop a bias towards the majority class, consequently failing to detect the crucial
but infrequent health events they were designed to find.
As such, the main goal of this PhD is to advance the practical application of ML
in health monitoring by developing and validating novel methodologies that directly
overcome these fundamental obstacles. The research aims to create a pathway
towards AI-driven systems that are not only accurate but also robust, data-efficient,
and equitable in their performance across diverse individuals and conditions.
To address these challenges, this dissertation introduces several key methodological
innovations. First, to combat poor model generalization, the work moves
beyond simplistic one-size-fits-all models, which often fail when applied to new
individuals with different genetics, behaviors, or physical builds. The contributions
in this area are multifaceted and include: i) developing personalized modeling
techniques, such as individual-specific calibration and dynamic anomaly detection
thresholds, which explicitly account for subject-specific physiological and
behavioral characteristics; ii) analyzing the impact of data transformations (e.g.,
the orientation-invariant L2-norm and time-frequency representations like the Continuous
Wavelet Transform) to create more robust data representations that are
less sensitive to inter-individual variations and sensor noise; and iii) performing an
SUMMARY xxv
in-depth investigation of advanced deep learning architectures, specifically Time-
Series Transformers, and rigorously evaluating their ability to generalize not only
across unseen patients but also across different data sampling rates, a critical factor
for real-world deployment where data is collected from various devices.
Second, to tackle the dual problems of data scarcity and class imbalance, which
severely hinder the application of standard supervised learning, this PhD presents a
suite of targeted solutions. The contributions designed to overcome these issues are:
i) the development of data-efficient algorithms, such as an approach for equine gait
classification that combines automated feature selection with clustering to replace a
data-hungry training phase with a brief, few-shot calibration, drastically reducing
the amount of labeled data required for accurate classification; ii) the application
of a fully unsupervised anomaly detection framework, utilizing a convolutional
autoencoder trained only on normal behavioral data, to entirely eliminate the need
for labeled examples of a difficult to observe event like foaling in mares; and iii)
the systematic implementation and evaluation of techniques such as weighted loss
functions, which penalize the misclassification of minority classes more heavily,
and the use of balanced performance metrics (e.g., balanced accuracy, class-specific
recall) to ensure models effectively learn from, and are fairly evaluated on, their
ability to detect these critical minority classes.
The results obtained throughout this PhD demonstrate the practical success and
tangible impact of these contributions. In response to data scarcity, the proposed
data-efficient algorithm for equine gait classification achieved an impressive 95%
accuracy using only approximately 200 seconds of labeled data for calibration, a
performance comparable to deep learning models requiring orders of magnitude
more data. Similarly, the unsupervised foaling detection framework successfully
identified 100% of parturition events in the study cohort, completely removing the
need for labeled examples of this difficult to observe event, albeit with a trade-off
in false alarms that could be tuned via a dynamic threshold. To address class
imbalance and generalization in human electrocardiogram (ECG) analysis, the
proposed methods yielded significant improvements; the systematic application of
automated feature selection and a weighted loss function improved the recall of the
rare supraventricular ectopic beat (SVEB) class by almost 20%, from a baseline
of 75.2% to a final recall of 94.2%. Furthermore, the investigation into advanced
architectures revealed that, without expert features, Time-Series Transformers
consistently outperformed convolutional neural networks (CNNs) by over 30% on
unseen patients and by over 20% on unseen sampling rates. The fusion of expert
knowledge in the form of RR-interval features further enhanced the generalization
of both architectures, pushing their performance to over 92% balanced accuracy on
unseen patients and sampling rates. Cumulatively, these results demonstrate that
the steps taken during this PhD form a strong and coherent foundation to support
the overarching goal of achieving more robust, data-efficient, and generalizable
AI-driven health monitoring systems, moving them closer to reliable, real-world
deployment.