UCL Discovery
UCL home » Library Services » Electronic resources » UCL Discovery

Screening for oral cancer and precancer in general dental practice: Evaluation of machine learning software in the identification of high risk groups

Lim, Kenneth; (2004) Screening for oral cancer and precancer in general dental practice: Evaluation of machine learning software in the identification of high risk groups. Doctoral thesis (Ph.D), UCL (University College London). Green open access

[thumbnail of Screening_for_oral_cancer_and_.pdf] Text

Download (13MB)


Background: There is much to recommend the early detection of oral cancer and precancer as this could lead to a reduction in mortality and morbidity, an increase in the quality of life, and cost savings for the Health Service. Oral cancer fulfils most of the criteria for a screenable disease; a simple visual examination is well accepted by patients and is a valid means for detecting oral cancer and precancer. However it is widely agreed that population screening for oral cancer and precancer cannot be recommended because of its relatively low prevalence in the United Kingdom. This study investigates the role of general dental practitioners in the detection of lesions, smoking and drinking habits within the population who visit their dentist. It also looks at the usefulness of machine learning software (artificial intelligence) for the identification of groups that could be at high risk of getting oral cancer or precancer. Methods: This study was carried out by 18 general dental practitioners from practices in north London, south London, Nottingham and Aldershot. Results: 2265 cases were available for analysis. Oral lesions were detected in 14.1% (319) of patients with 4.2% (94) of these considered to be positive. This included 2 cases of squamous cell carcinoma one developing from a case of oral submucous fibrosis. The prevalence of white patches was 2.0%, red patches, 0.5% and lichen planus was 1.5%. In this study, the prevalence rate for positive lesions and conditions was 4.2%. 27% of the screened population were smokers and 16.5% consumed more than 5 units of alcohol a week. There were significant correlations between positive lesions and male gender (IRR 1.86, 95%C1 1.22-2.82), heavy smoking males (IRR 3.68, 95%CI 2.10-6.43), heavy smoking female (IRR 3.58, 95% CIl.35-9.50) and heavy alcohol use in males (IRR 2.98, 95% Cl 1.06-3.47) Machine learning models performed well and were capable of achieving high sensitivities of 85% and 80% however always accompanied by low specificities of 27% and 31%. Different types of machine learning methods were used and none performed better than another. Data from the 1995 studies showed remarkable epidemiological similarities with data from this current study despite the interval between the two projects being some 5 years and studies done in different environments, though all mainly in the London area. Conclusions: The results suggested that general dental practitioners were able to identify oral mucosal lesions and conditions following screening according to criteria that they had been taught. The prevalence of risk factors such as smoking and drinking habits within the screened group were similar to those reported in national statistics; the prevalence of lesions and conditions regarded as positive were also comparable to published studies in the UK and other countries. This suggests that the population visiting general dental practices is representative of the general population and that screening in general dental practice is feasible; also that the general dental practice environment appears well suited to opportunistic screening for oral cancer and precancer as about 60% of the population regularly attend at their dentist. Thus the opportunity may exist for screening a large proportion of the general population as part of routine dental patient care, in a cost-effective manner. Machine learning models performed well in identifying positive cases but with unacceptably high numbers of false positives. At this stage of development machine learning models cannot be recommended for use as a pre-screen filter although it could have a future role as part of an interactive health awareness package to patients. Building on knowledge and experiences gathered from this study, further screening studies are envisaged involving dental practices in other regions to explore possible variations in prevalence of oral cancer and precancer within the UK. (Abstract shortened by UMI.)

Type: Thesis (Doctoral)
Qualification: Ph.D
Title: Screening for oral cancer and precancer in general dental practice: Evaluation of machine learning software in the identification of high risk groups
Open access status: An open access version is available from UCL Discovery
Language: English
Additional information: Thesis digitised by ProQuest.
URI: https://discovery.ucl.ac.uk/id/eprint/10100464
Downloads since deposit
Download activity - last month
Download activity - last 12 months
Downloads by country - last 12 months

Archive Staff Only

View Item View Item