(figure 5.2)
1.0
0.8 0.6area = 0.84
0.4
0.2 0.00.0
0.2
0.4
0.6
0.8
1.0Table 5.4 Specialist result compared with neural network’s screening result
Specialist
Neural Network Positive Negative Total
Positive 8 82 90
Negative 2 273 275
Total 10. 355 365
Sensitivity: (8/10) = 0.80 Specificity: (273/355) = 0.77 Positive predictive value: (8/90) = 0.09 Negative predictive value: (273/275) = 0.99 Positive prevalence: (10/365) = 2.7% Likelihood ratio (0.8/0.23) =3.48
Table 5.5 Specialist result compared with dentists’ screening result.
Specialist
Screener Positive Negative Total
Positive 40 20 60
Negative 16 1951 1967
Total 56 1971 2027
Sensitivity: (40/56) = 0.71 Specificity: (1951/1971) = 0.99 Positive predictive value: (40/60) = 0.67 Negative predictive value: (1951/1967) = 0.99 Positive prevalence: (56/2027) = 2.8% Likelihood ratio (0.74/0.01) = 74.00
5.5 Discussion
Logistic regression analysis is widely used throughout epidemiology to calculate odds ratios. The odds ratio is used to relate disease risk to exposure. Brugere et al,
(1986), by use of this method, demonstrated a relationship between the site of oral cancer and level of daily alcohol consumption whilst Mashberg et al (1993) were able to demonstrate an increasing odds ratio of oral cancer with tobacco (35 cigarettes a day) and alcohol (21 whisky equivalents a day). By evaluating the relative risk of different lifestyle habits it is possible to use this information for developing more sophisticated means of identifying individuals who could be targeted for screening or health education. Wilkinson et al (1994) evaluated a risk scoring system for cervical cancer to be used in primary care. The system enabled identification of 75 % of those women with cervical neoplasia, they concluded that further research was required to assess the effectiveness of risk targeting.
Although logistic regression is a useful tool for assessing risks it would not be practical in the primary care setting, where time and ease of use is of the essence. The use of neural networks, which has become more accepted in recent years, may be more feasible. There are few references to neural networks in the literature prior to 1988 but their use is increasing in the field of cancer control, prevention and treatment.
In the screening programme reported in Chapter 3 the overall performance of the dentists was superior to that of the neural network. However the sensitivity of the
neural network was similar to that of the junior hospital dentists. This is not surprising since the network will have determined that most of the subjects with lesions were smokers and/ or drinkers. The specificity achieved by the neural network was, however quite low with a false positive rate of 23%, and the odds of having a lesion if classified as positive by the network were only 3.5 compared to over 60 if screened positive by a dentist. This was probably due to the network selecting all those individuals who, from their risk habits could be considered to have a high likelihood of being diagnosed positive but had not developed lesions. In a preliminary screening procedure such as this, a high false positive rate is not a cause for concern, and indeed may be beneficial since these individuals will be subjected to a further test (oral examination) and can be selected for preventive education. In a similar study in which a neural network was used to predict breast cancer malignancy Floyd et al
(1994) found that the network achieved sensitivity of 1.0 and a specificity of 0.59. Ercal et al (1994) found that their neural network designed to detect malignant melanomas from colour images was able to classify correctly over 80% of the malignant and benign tumours on real skin images. Snow et al (1994) used a neural network to assess the diagnosis and prognosis of prostate cancer. The network was able to predict a biopsy result with 87 % overall accuracy from serum prostate specific antigen levels. It was also able to predict tumour recurrence with 90% overall accuracy.
The role of neural networks in screening programmes may be as an adjunct in identifying high risk individuals. If the cost of setting up a screening programme and the cost of a dentist’s time is taken into account, then a neural network may prove to
be more economical since it could make an a priori selection of high risk individuals that ought to be screened by a clinician. In dental practice, the system could be used to assign a risk status to a patient in order to help decide who should receive a detailed oral mucosal examination.
Although artificial intelligence is relatively new to the field of medicine and dentistry, its usefulness in clinical decision making is becoming more apparent. This part of the study has shown that this system, or a more user friendly version, could have a place in the dental surgery. A simulation of how a neural network (NN) would be able to predict the likelihood of an individual having a precancerous or cancerous lesion of the oral mucosa, given the age, gender, smoking, drinking and dental habits of each screened individual, is shown in Figure 5.3. In the economic climate which exists today it is important to get the best output from the minimum input in terms of cost and patient satisfaction. If the neural network is used as a filtering mechanism it avoids the unnecessary cost, time and potential distress of screening low risk asymptomatic patients.
Despite the projected use of advanced computer technology in the early detection of oral cancer, it is still of utmost importance to increase awareness of oral cancer and its risk factors within the dental and medical profession and general public. From the findings in this study it is tentatively proposed that neural networks will have an increasing place in the early detection of oral cancer following further studies.