Modelling of vital-sign data from post-operative patients
<p>Thousands of in-hospital deaths each year in the UK are potentially preventable, being often preceded by physiological deterioration. The current standard of clinical practice for patient monitoring on general wards is the periodic observation of vital signs by nursing staff. The use of ear...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
2015
|
| Subjects: |
| _version_ | 1797063074758262784 |
|---|---|
| author | Pimentel, M |
| author2 | Tarassenko, L |
| author_facet | Tarassenko, L Pimentel, M |
| author_sort | Pimentel, M |
| collection | OXFORD |
| description | <p>Thousands of in-hospital deaths each year in the UK are potentially preventable, being often preceded by physiological deterioration. The current standard of clinical practice for patient monitoring on general wards is the periodic observation of vital signs by nursing staff. The use of early warning score (EWS) systems should enable a more timely response to, and assessment of, acutely ill patients. The investigations described in this thesis seek to apply principled approaches based on machine learning to the analysis of vital-sign data from patients who are recovering from major surgery.</p> <p>A dataset comprising observational vital-sign data from 407 post-operative patients taking part in a two-phase clinical trial in the Oxford Cancer Centre is introduced. A second independent dataset collected from clinical data obtained from 24,212 patients admitted to the Medical Assessment Unit of a different hospital is used for validation purposes. When applied to post-operative patients, currently-used EWS systems achieve values of Area Under the Receiver-Operating Characteristic curve (AUROC) that range from 0.717 to 0.841 for predicting a composite outcome of death, emergency admission to the Intensive Care Unit, and cardiac arrest within 24 hours. We also demonstrate that the method of recording vital signs on the ward plays a fundamental role in the design and performance of EWS systems. Using the same set of physiological variables, kernel density estimators and support vector machines give equivalent results to those of EWS systems which have been carefully optimised by trial and error.</p> <p>A method for describing the physiological trajectories of post-operative patients is developed using machine learning techniques. We further introduce the concept of variability of vital signs over a 24-hour period, and propose a strategy for incorporating this information into the machine learning models studied. The resulting model leads to an improvement in performance (AUROC = 0.856). An approach based on Gaussian processes is then discussed for exploring and representing patterns of vital-sign time-series data. The approach allows different types of normal physiological trends to be identified in patients recovering from surgery.</p> <p>Knowledge of different patterns among hospitalised patients and their incorporation in monitoring systems improves early-warning scoring systems for the identification of physiological deterioration in specific patient groups.</p> |
| first_indexed | 2024-03-06T20:54:40Z |
| format | Thesis |
| id | oxford-uuid:38c7598e-944f-4e2d-8c9f-4c902ee3a89c |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-03-06T20:54:40Z |
| publishDate | 2015 |
| record_format | dspace |
| spelling | oxford-uuid:38c7598e-944f-4e2d-8c9f-4c902ee3a89c2022-03-26T13:52:05ZModelling of vital-sign data from post-operative patientsThesishttp://purl.org/coar/resource_type/c_db06uuid:38c7598e-944f-4e2d-8c9f-4c902ee3a89cPatient monitoringMachine learningPatient monitoring--Data processingSignal processingEnglishORA Deposit2015Pimentel, MTarassenko, LYoung, DClifton, DWilliams, C<p>Thousands of in-hospital deaths each year in the UK are potentially preventable, being often preceded by physiological deterioration. The current standard of clinical practice for patient monitoring on general wards is the periodic observation of vital signs by nursing staff. The use of early warning score (EWS) systems should enable a more timely response to, and assessment of, acutely ill patients. The investigations described in this thesis seek to apply principled approaches based on machine learning to the analysis of vital-sign data from patients who are recovering from major surgery.</p> <p>A dataset comprising observational vital-sign data from 407 post-operative patients taking part in a two-phase clinical trial in the Oxford Cancer Centre is introduced. A second independent dataset collected from clinical data obtained from 24,212 patients admitted to the Medical Assessment Unit of a different hospital is used for validation purposes. When applied to post-operative patients, currently-used EWS systems achieve values of Area Under the Receiver-Operating Characteristic curve (AUROC) that range from 0.717 to 0.841 for predicting a composite outcome of death, emergency admission to the Intensive Care Unit, and cardiac arrest within 24 hours. We also demonstrate that the method of recording vital signs on the ward plays a fundamental role in the design and performance of EWS systems. Using the same set of physiological variables, kernel density estimators and support vector machines give equivalent results to those of EWS systems which have been carefully optimised by trial and error.</p> <p>A method for describing the physiological trajectories of post-operative patients is developed using machine learning techniques. We further introduce the concept of variability of vital signs over a 24-hour period, and propose a strategy for incorporating this information into the machine learning models studied. The resulting model leads to an improvement in performance (AUROC = 0.856). An approach based on Gaussian processes is then discussed for exploring and representing patterns of vital-sign time-series data. The approach allows different types of normal physiological trends to be identified in patients recovering from surgery.</p> <p>Knowledge of different patterns among hospitalised patients and their incorporation in monitoring systems improves early-warning scoring systems for the identification of physiological deterioration in specific patient groups.</p> |
| spellingShingle | Patient monitoring Machine learning Patient monitoring--Data processing Signal processing Pimentel, M Modelling of vital-sign data from post-operative patients |
| title | Modelling of vital-sign data from post-operative patients |
| title_full | Modelling of vital-sign data from post-operative patients |
| title_fullStr | Modelling of vital-sign data from post-operative patients |
| title_full_unstemmed | Modelling of vital-sign data from post-operative patients |
| title_short | Modelling of vital-sign data from post-operative patients |
| title_sort | modelling of vital sign data from post operative patients |
| topic | Patient monitoring Machine learning Patient monitoring--Data processing Signal processing |
| work_keys_str_mv | AT pimentelm modellingofvitalsigndatafrompostoperativepatients |