Who this list is for
Edge computing means processing data on or near the device that produces it, such as a wearable sensor, a bedside monitor or a server inside the hospital, instead of sending everything to a distant cloud. For clinical AI this can shorten response times, keep working when connections fail and reduce the amount of identifiable data that leaves a site. It also brings new problems: small devices have limited memory and power, models must be updated safely, and every device becomes something to secure.
This list is for doctoral researchers in health informatics, biomedical engineering and computer science who want a topic in this space. The result is six topics in four themes.

How to use this list
Read each row as a starting point. Pick one clinical problem first and the technology second: a thesis on "edge AI for healthcare" in general has no clear population, outcome or comparison. The columns suggest a research question, the likely design and data, and the reason the topic matters now. If you are still choosing a broad area, start with how to choose a PhD topic. Related lists: AI in brain disorder diagnosis and digital therapeutics in connected healthcare.
Theme 1: Real-time diagnostics and remote monitoring
The strongest case for edge AI is where seconds matter or connectivity is unreliable: deterioration on a ward, arrhythmia alerts, or monitoring at home. The research gap is less about whether models can run on small hardware and more about whether they keep their accuracy when compressed, and whether their alerts help clinicians. Reporting guidance for clinical prediction models now expects authors to describe the model, its data and its evaluation in detail (Collins et al., 2024), and early clinical evaluation of AI decision support should examine how clinicians use the output in practice (Vasey et al., 2022).
| Topic | Possible research question | Likely design and data | Why it matters now |
|---|---|---|---|
| On-device early warning of clinical deterioration: accuracy and latency compared with a cloud model | Does a compressed early-warning model running on bedside hardware match the discrimination and calibration of its full cloud version while reducing alert latency? | Retrospective development on routinely collected vital-sign data, model compression, bench testing of latency and energy, then a silent prospective run | Clinical AI must be reported and evaluated transparently before deployment (Collins et al., 2024) |
| Edge-based remote monitoring to predict exacerbations in one chronic condition | Can on-device processing of home sensor data predict exacerbations of a chosen condition (for example heart failure or COPD) early enough to support triage? | Prospective observational cohort with home devices, or secondary analysis of an existing monitoring dataset; time-to-event or prediction metrics; clinician interviews on alert usefulness | Early clinical evaluation should test how AI output changes decisions, not only accuracy (Vasey et al., 2022) |
Keep the clinical comparison explicit: the question is not whether an edge model works, but whether it works as well as the current alternative for a defined patient group.
Theme 2: Wearables and rehabilitation
Wearable sensors produce continuous data that are costly to transmit and store, which makes on-device feature extraction attractive. In a systematic review of 32 studies of wearable sensors with machine learning in rehabilitation training, Wei and Wu (2023) found that many relied on simulated rather than real patients, while Daniore et al. (2024) drew ten lessons from wearable studies in multiple sclerosis on turning sensor data into digital biomarkers, starting with aligning the device with the research question. A doctoral project can test whether edge processing preserves the clinical validity of such measures.
| Topic | Possible research question | Likely design and data | Why it matters now |
|---|---|---|---|
| On-device processing of wearable data for personalised rehabilitation monitoring | Does on-device feature extraction from a wearable sensor give gait or activity measures that agree with laboratory reference measures in patients after stroke or joint replacement? | Validation study against a reference system, agreement statistics, battery and data-volume measurement; small feasibility study of patient use | Many wearable machine learning studies in rehabilitation used simulated rather than real patients (Wei & Wu, 2023; Daniore et al., 2024) |
Theme 3: Hospital automation and robotics
Surgical robotics is regulated as a medical device and needs specialist facilities, so it rarely suits a single doctorate. A more feasible route is one non-surgical task, such as a ward robot that delivers supplies or detects falls, where edge processing affects response time and safety.
| Topic | Possible research question | Likely design and data | Why it matters now |
|---|---|---|---|
| Edge-hosted perception for a non-surgical hospital robot: response time, safety and staff acceptance | Does running perception on the robot or a local server, rather than in the cloud, improve response time and reduce unsafe events in a defined ward task, and how do staff judge it? | Simulation and controlled trials in a test ward, logged response and incident data, staff questionnaires and interviews | Digital health technologies are expected to show evidence proportionate to their risk (National Institute for Health and Care Excellence, 2022) |
Theme 4: Federated learning, privacy and security
Federated learning trains a shared model without pooling patient records: each hospital or device trains locally and only model updates are combined. Kaissis et al. (2020) set out how federated and other privacy-preserving methods could support medical imaging AI, and Teo et al. (2024) systematically review clinical applications and architectures, finding that only about one study in twenty described real-life application. Privacy is not automatic: Nasr et al. (2019) show that model updates in federated learning can be attacked to infer information about training data.
| Topic | Possible research question | Likely design and data | Why it matters now |
|---|---|---|---|
| Federated learning across hospitals for one diagnostic task: performance against centralised training | How close does a federated model come to a centrally trained model when data differ between sites, and which aggregation strategy narrows the gap? | Multi-site or simulated multi-site experiment on a public or consortium dataset; site-level performance and calibration; heterogeneity analysis | Only about one in twenty federated learning studies in health care described real-life use (Teo et al., 2024) |
| Privacy and security risks of edge AI deployments in a hospital network | Which attacks on edge devices and federated updates are most feasible in a hospital network, and how much do defences cost in accuracy and latency? | Threat modelling, controlled attack experiments in a testbed, trade-off analysis of defences | Federated updates can leak information, so defences need testing in realistic settings (Nasr et al., 2019; Kaissis et al., 2020) |
How to test a topic before you commit
Run each candidate through the FINER criteria: is it feasible, interesting, novel, ethical and relevant (Hulley et al., 2013)? For edge AI topics, four checks matter most.
- Data: confirm you can obtain clinical data for development and a separate set for testing. Public datasets help with development but rarely reflect your target hardware or setting.
- Hardware: decide early which devices you will use and measure latency, memory and energy on them.
- Clinical partner: a clinician who owns the use case keeps the outcome measures relevant.
- Ethics and regulation: prospective work with patients needs ethics approval; software that informs diagnosis may fall under medical device rules, so plan its evidence accordingly (National Institute for Health and Care Excellence, 2022).
For a step-by-step method see how to select a PhD topic and the topic selection service page.
Common mistakes
- Naming a technology ("edge AI framework") without a patient group, outcome or comparison.
- Reporting only accuracy and ignoring calibration, latency, energy and failure modes on real hardware.
- Claiming a system is "secure" or "private" without a threat model and tests.
- Bundling robotics, wearables, federated learning and security into one thesis.
- Citing papers from unrelated fields to fill a reference list; read every source you cite.
Key takeaways
- Start from one clinical use case and compare the edge system with the current alternative.
- Report clinical AI models transparently and evaluate them in real workflows (Collins et al., 2024; Vasey et al., 2022).
- Wearable and rehabilitation studies need careful validation against reference measures (Wei & Wu, 2023; Daniore et al., 2024).
- Federated learning is rarely tested in real clinical use, and it is not private by default (Teo et al., 2024; Nasr et al., 2019).
- Test feasibility, data access, ethics and regulation before you commit (Hulley et al., 2013).
