University of Turku Health Technology Master’s Thesis Guide: DTEK2066, 30 ECTS, Biosignals, Machine Learning and UTUGradu

Quick answer: what is the University of Turku Health Technology thesis route?

The current Master’s Degree Programme in Health Technology is a 120 ECTS, two-year Master of Science (Technology) programme in the Faculty of Technology. The controlling curriculum is HEALTHTECH2427 / Peppi programme 99379 for 2024–2027. Its “Master’s thesis and project” module is 40 ECTS, but the actual thesis is DTEK2066 Master’s Thesis in Technology, Biomedical Engineering and Health Technology, 30 ECTS. The remaining module consists of TTDK1308 Degree Qualifying Examination, 0 ECTS, plus one 10 ECTS project choice: DTEK0088 Capstone or DTEK2056 Lean Digital Business Design. Do not call the thesis itself 40 ECTS.

1. Read the 120 ECTS degree without double-counting

The current programme structure contains 20 ECTS Common Studies, 20 ECTS Health Technology core studies, 0–5 ECTS language studies, the 40 ECTS thesis-and-project module, 20–25 ECTS thematic/minor studies and 10–20 ECTS other studies. Peppi’s technical ranges allow choices, but the programme degree is 120 ECTS. Use the current personal study plan to resolve the final combination.

2. The actual thesis is DTEK2066, 30 ECTS

DTEK2066 is Advanced Studies in Biomedical Engineering and Health Technology. It carries 30 ECTS, is graded 0–5, and can be completed in Finnish or English. The course expects scientific work, command of research methods, subject knowledge and scientific writing. The intended outcome is a constructive solution proposal to a research challenge, often from a company, organisation or research group.

3. The other 10 ECTS is a project, not thesis credit

Current Peppi requires one 10 ECTS project alternative inside the 40 ECTS module. Students choose DTEK0088 Capstone or DTEK2056 Lean Digital Business Design. These courses build engineering, teamwork, prototyping or business-development competence. They do not change DTEK2066 into a 40 ECTS thesis and they do not replace its scientific examination.

4. Choose Capstone or Lean Digital Business Design deliberately

DTEK0088 follows an engineering conceive-design-implement-operate process for open-ended real-world challenges. DTEK2056 develops digital products/services and business models through lean-startup, customer-development and prototyping work. They are alternatives. Select the one that fits the degree plan and skill goals; do not count both as compulsory thesis components.

5. Start DTEK2066 from an accepted research plan

The programme page states that thesis work is performed according to an accepted research plan, independently but under supervision. DTEK2066 says the student and assigned supervisors create the plan at the beginning and update it when necessary. Make the plan specific enough to review the question, data, measurement, method, validation strategy, risks and timeline before the main work becomes expensive to change.

6. A company topic still needs a research question

Health Technology encourages company-supplied topics, but a product task is not automatically a thesis question. Convert “build an ECG classifier” into an academically evaluable problem: for which population, signal conditions, target label and deployment scenario, and against what baseline? The thesis must explain not only what was built but what was learned through defensible scientific analysis.

7. Supervision is regular, not only at the beginning and end

DTEK2066 expects progress reporting to supervisors in agreed forms. Use those checkpoints to confirm that data access, labels, sensor setup, preprocessing and evaluation still support the research question. If the plan changes, record the decision and its reason rather than rewriting history in the final manuscript.

8. The examiner chain is programme-specific and explicit

The current DTEK2066 course says the thesis is evaluated by at least two examiners under University grading guidelines. The final acceptance is decided by the head of the department, and the grade is based on examiner evaluation. This is a current course-specific rule and should not be replaced with the Food Development examiner process or another Faculty of Technology programme’s practice.

9. The seminar exists, but do not invent seminar credits

DTEK2066 states that thesis achievements and process are presented in a related Master’s thesis seminar. A current UTU Moodle realization exists for DTEK2067 Master’s Thesis Seminar, Biomedical Engineering and Health Technology. However, current Peppi programme 99379 does not allocate DTEK2067 any credits in its 40 ECTS thesis/project module. Follow the current seminar instructions, but do not add a made-up 5 ECTS seminar or alter the thesis credit unless your current HOPS/Peppi explicitly says so.

10. TTDK1308 is 0 ECTS but remains part of the route

The current module includes TTDK1308 Degree Qualifying Examination for Master’s Degree, 0 ECTS. Normally the thesis abstract or another suitable part functions as the maturity test and demonstrates familiarity with the thesis field. A separate written test can apply in the specific language-education circumstances described by the current course.

11. Health Technology is a measurement-and-data programme

The programme focuses on health-data measurement and analysis, signal processing, machine learning, sensor technology and Internet of Things. Typical applications include wellness systems, diagnostic technologies and remote monitoring. A thesis can therefore sit at very different layers: sensor hardware, embedded software, signal processing, prediction models, validation or clinical-facing analytics.

12. Define the evidence layer before choosing the metric

A sensor-prototype thesis, an algorithm thesis and a clinical-prediction thesis do not have the same success criterion. Hardware may need signal fidelity and power tests. Signal processing may need event-detection error or noise robustness. Prediction work may need patient-level generalisation. Clinical-facing work may need sensitivity, specificity, calibration or prospective validation. Match the metric to the intended claim.

13. Biosignal acquisition is part of the method, not background detail

DTEK0042 covers ECG, EEG, EMG and PPG, including signal origin, sensor technology, acquisition, preprocessing, feature extraction and vital-sign detection. Record sensor model, placement, sampling characteristics, recording protocol and relevant synchronization. A model cannot recover physiological information that acquisition never captured reliably.

14. Preprocessing can change the result

Filters, resampling, artefact rejection, normalization and segmentation can alter waveform morphology and downstream features. Specify these operations and their parameters. If preprocessing differs by group or is tuned after seeing model performance, document that choice. Preserve a reproducible route from raw recording to the final input matrix.

15. A signal window is not automatically a new patient

One person can contribute thousands of ECG beats, PPG windows or EEG segments. Those windows share subject-level characteristics. If the research question concerns new patients, random window-level train/test splitting can leak patient identity and make performance look better than it will be on genuinely unseen patients.

16. Decide the validation unit before model development

TKO_7092 explicitly distinguishes prediction for new measurements from known patients from prediction for entirely new patients. Ask what the real deployment scenario is. If deployment is to new people, split at patient level. If deployment is future windows from an already enrolled patient, a different evaluation design may be justified. State the estimand before selecting the resampling scheme.

17. Use nested evaluation when tuning and assessment must be separated

Cross-validation can support model selection, but performance estimation becomes optimistic if hyperparameters, features or thresholds are repeatedly optimized against the same evaluation folds. TKO_7092 covers nested cross-validation for separating selection from assessment. The exact design should follow the research question and sample structure, not a ritual use of a particular number of folds.

18. Keep the final test set genuinely final

If a hold-out set is intended as an unbiased final estimate, do not use it to choose preprocessing, features, architecture, threshold or early stopping. Each look leaks information into development decisions. Keep a development record so the final reported metric comes from a model whose important choices were made without consulting the final test labels.

19. Accuracy is often inadequate for health data

Class imbalance is common in disease detection and monitoring. A model can have high accuracy while missing most positive cases. Depending on the task, report sensitivity, specificity, precision, recall, predictive values, ROC/PR measures, calibration, error distributions or clinically meaningful thresholds. Explain the denominator and prevalence context.

20. Internal validation is not external clinical validation

Cross-validation estimates performance under the sampled data-generating structure. It does not prove performance on another hospital, device, demographic group or disease spectrum. Device hardware, recruitment, prevalence, labelling and clinical workflow can all shift. If external validation is absent, say so and keep the conclusion within the evaluated domain.

21. Signal quality belongs inside the pipeline

DTEK2095 explicitly covers biosignal-quality estimation, preprocessing, peak detection, HR/HRV estimation, rhythm-disorder detection, outlier handling and post-processing. Do not quietly discard difficult signals and then call the algorithm robust. Report how much data was excluded, why, and what performance looks like under realistic quality conditions when possible.

22. Detection errors propagate to derived physiological variables

R-peak or pulse-peak errors change heart rate, inter-beat intervals and HRV. Evaluate the event detector before interpreting derived variables. If a result depends on post-processing that removes implausible intervals or repairs missed beats, document that stage. Derived physiology is only as trustworthy as the upstream detection pipeline.

23. Wearable hardware needs version control

DTEK2094 covers sensors, microcontrollers, embedded firmware, communication protocols, Bluetooth Low Energy and energy constraints. Record firmware version, sensor configuration, sampling settings and device revision. A silent firmware change can alter filtering, timestamps or missingness and make data from nominally identical devices non-equivalent.

24. Time synchronization can become the hidden failure mode

Connected health systems may combine multiple sensors, phones or reference devices. Clock drift, packet loss and asynchronous sampling can corrupt feature alignment. Define how streams are synchronized and what happens when timestamps disagree. For sensor-fusion work, calibration and timing files are part of the scientific record.

25. A working prototype is not a validated medical device

A wearable that measures a signal and transmits it to a phone demonstrates engineering feasibility. It does not automatically establish analytical validity, clinical performance, safety, usability or regulatory readiness. State which validation stage the thesis actually covers. Avoid medical-device claims that outrun the study.

26. Deep learning does not remove the need for study design

TKO_7094 covers training, regularization, CNNs, transfer learning and sequential data. Larger models can fit complex biosignals, but they can also exploit nuisance structure, site identity or patient-specific patterns. Regularization helps optimization and generalisation, but it does not fix leakage or a validation scheme that answers the wrong question.

27. Explainability is not proof of mechanism

Feature importance, saliency or attention can help inspect a model, but an explanation map is not automatically a biological causal mechanism. Validate whether the highlighted pattern is stable and clinically plausible, and separate “the model used this information” from “this physiological mechanism caused the disease.”

28. Computer vision and sensor fusion add calibration dependencies

TKO_7096 covers image classification, detection, segmentation, tracking, camera calibration and sensor fusion. In health-facing image or motion work, record calibration, spatial alignment and data provenance. If several images or frames come from one patient, patient-level splitting remains important even when the model operates on individual images.

29. Statistical inference must respect dependency

TKO_7093 covers exploratory analysis, statistical testing and method selection. Choose the independent unit based on study design, not spreadsheet row count. Repeated visits, signal windows or bilateral measurements can require repeated-measures or hierarchical treatment. Report missingness, exclusions and derived labels clearly.

30. Technical performance and clinical utility are different claims

A classifier can discriminate cases in retrospective data without improving a clinical decision. A wearable can estimate heart rate without being appropriate for diagnosis. A statistically better model may not produce a meaningful workflow benefit. Separate technical feasibility, analytical performance, clinical performance and clinical utility in the Discussion.

31. Prospective and retrospective evidence are not interchangeable

Retrospective datasets are efficient for algorithm development, but labels and measurement conditions may reflect past workflows. Prospective studies can reveal missingness, device failure, clinician interaction and population shifts that retrospective evaluation misses. If the thesis is retrospective, do not describe the result as prospective clinical effectiveness.

32. Plan ethics from the actual study, not the programme name

Human participants, clinical records, wearable monitoring and hospital collaboration can create ethics obligations, but the route depends on the design. University guidance distinguishes relevant ethics pathways and requires review before data collection when review is necessary. Resolve consent, data access, ethics and host permissions before collecting or receiving restricted data.

33. A research permit does not replace ethics or consent

A hospital, company or other data-holding organisation may require a research permit. That authorizes work within that organisation’s governance; it does not automatically satisfy participant consent or ethical-review requirements. Treat these as separate gates and keep the documentation with the project record.

34. Health data can remain identifiable after names are removed

ECG, EEG, PPG, device identifiers, timestamps, rare diagnoses and demographic combinations can create re-identification risk. Pseudonymised data are still personal data when a link or reasonable re-identification route remains. Minimise data, control access and avoid exporting protected records into unapproved tools.

35. Company confidentiality must be planned before writing

Accepted University of Turku theses are public documents, and confidential information should not be included. Company-sponsored projects should separate protected source material from what the public thesis needs to disclose. Resolve proprietary algorithms, datasets, device internals and unpublished results early, not when the final PDF is already complete.

36. Turnitin is an originality control, not a model audit

Degree theses undergo mandatory Turnitin checking, with Master’s approval through UTUGradu. Similarity checking cannot tell whether patient-level splitting was correct, whether a filter distorted a biosignal or whether calibration is valid. Scientific validity, originality and privacy are different controls and all must be satisfied.

37. UTUGradu is the institutional submission route

UTUGradu manages originality checking, examination, approval, publication and archiving for higher-degree theses. Before submission, freeze the exact final manuscript, verify metadata, remove confidential or identifiable content, and ensure figures and tables match the final analysis. Keep the submitted version distinct from working drafts.

38. Use the examiner route as a final scientific stress test

Because at least two examiners assess DTEK2066 and the department head decides final acceptance, make the manuscript auditable by someone who did not build the system. Define the data-generating process, preprocessing, validation unit, baselines, uncertainty and limitations explicitly. Reproducibility should not depend on undocumented oral knowledge from the research group.

39. A practical 12-step Health Technology workflow

1. Confirm programme 99379 and the current HOPS. 2. Select DTEK0088 or DTEK2056 for the separate 10 ECTS project. 3. Define a thesis question and supervisors. 4. Build and obtain acceptance for the research plan. 5. Resolve data access, ethics, permits and confidentiality. 6. Freeze acquisition/preprocessing and the independent validation unit. 7. Collect or receive data with traceable subject/session identifiers. 8. Build models or engineering systems only inside the development partition. 9. Evaluate with deployment-relevant validation. 10. Present progress through the current thesis-seminar process without inventing credits. 11. Complete TTDK1308, Turnitin and DTEK2066 examination. 12. Submit the exact public-safe final version through UTUGradu.

40. Final pre-submission checklist

Confirm that DTEK2066 is still the current 30 ECTS thesis, the 10 ECTS project choice is separately complete, TTDK1308 is handled, and any seminar requirement is taken from current HOPS/instructions rather than assigned unsupported credits. Check supervisor-approved plan, raw-to-result provenance, patient-level independence where appropriate, no leakage from final test data, realistic signal-quality handling, uncertainty and limitations, ethics/permissions, confidentiality, Turnitin, the two-examiner route, department-head acceptance process and the exact UTUGradu file.

Final reproducibility check

Re-run at least one headline result from the frozen source data and final analysis configuration. Confirm subject counts, label definitions, preprocessing, split membership, model version, metric and figure/table value. If the number cannot be regenerated without manual undocumented edits, the pipeline is not yet ready for examination. For an engineering thesis, this spot check is often the fastest way to detect stale plots, accidental leakage or results from an earlier device/software version.

Additional execution notes for health-data provenance

Build a subject/session manifest before analysis. Each recording should connect a study subject identifier, visit or session, device, firmware or acquisition configuration, start/end time, raw-file path, label source and any quality flag. Keep personally identifying keys outside the analysis table under the approved data-governance arrangement. The analysis identifier should remain stable even if a file is renamed. This simple manifest prevents a common failure where repeated recordings from one person accidentally cross validation boundaries or labels are attached to the wrong signal file.

Treat reference labels as measurements with their own uncertainty. A diagnosis taken from a clinical record, an expert ECG annotation, a device-generated rhythm label and a questionnaire answer are not interchangeable ground truths. Document who or what generated the label, at what time relative to the signal, whether adjudication occurred, and how disagreement or ambiguous cases were handled. If the target is derived by an algorithm, the thesis should not describe it as an independent clinical truth without qualification.

For acquisition or device-comparison studies, define the reference system and synchronization procedure. If a wearable PPG device is evaluated against ECG, explain how the streams are time-aligned and what tolerance is accepted when pairing events. If a sensor is compared with a laboratory reference, report the calibration range and conditions. Agreement, correlation and classification accuracy answer different questions. High correlation can coexist with clinically important systematic bias, so the statistic should match the measurement claim.

Additional execution notes for leakage and validation audits

Before the final model run, create a leakage checklist. Verify that no patient appears in both development and final-test partitions when the target claim is new-patient generalisation. Confirm that normalization parameters, imputation values, feature selection and dimensionality reduction are fitted using training data only inside each resampling loop. Check whether augmented samples, overlapping windows or duplicate recordings derived from the same source can cross folds. Inspect whether filenames, site codes, acquisition devices or timestamps accidentally encode the outcome.

If the dataset includes multiple hospitals, devices or collection periods, use them to probe distribution shift when sample size permits. A temporal split can answer whether a model survives later data. A device-held-out analysis can reveal dependence on hardware. A site-held-out analysis can be more informative than random cross-validation for deployment elsewhere. These analyses are not always mandatory, but the thesis should explain which shift it can and cannot evaluate.

Report the full denominator behind every headline metric. State how many participants, sessions and windows entered development, validation and final evaluation, how many were excluded and why, and whether class prevalence differs across partitions. For confidence intervals or repeated resampling, describe what is being resampled. Reporting thousands of signal windows without the underlying participant count can make a small study appear much larger than it is.

Additional execution notes for software and model provenance

Freeze the code revision, environment and model configuration that generated the final result. Record important library versions, random seeds where relevant, trained-model checksum or identifier, and configuration files. Do not overwrite the best model every time an experiment runs. A lightweight experiment table with data version, split version, model parameters and output path is enough to make the thesis auditable without requiring enterprise infrastructure.

For deep models, separate architecture selection, training stopping criteria and threshold selection from final evaluation. If pretrained weights are used, identify their source and intended domain. Transfer learning from general images or unrelated signals can be useful, but it does not guarantee clinically meaningful representation. If an external model or dataset has usage restrictions, confirm that the thesis can lawfully reproduce or describe the work before publication.

Additional execution notes for clinically cautious conclusions

Write conclusions at the strongest level directly supported by the study. A retrospective patient-level evaluation may support “the model distinguished X from Y in this dataset under this validation scheme.” It does not automatically support “the system diagnoses X” or “the model improves care.” A prototype study may support feasibility. A measurement-comparison study may support agreement within tested conditions. A prospective clinical-utility claim generally needs evidence beyond technical algorithm performance.

Before UTUGradu submission, ask a reader outside the project to trace one result from the research question to raw data, preprocessing, split, model, metric and conclusion. If any step depends on an undocumented assumption, add it to Methods or limitations. Health Technology theses often combine engineering and medical language, so this external read-through is especially useful for catching claims that sound clinically stronger than the actual technical evidence.