STA-21 br Fig Number of FA
Fig. 4. Number of FA with sensitivity (in %) for various enumeration of budget.
thus becoming the best option in this enumeration list. As stated earlier in Section 4.1 we find the same tendency of FA becoming high for the combination of tests with lower sensitivity. This is to make the threshold high to achieve 100% sensitivity. For example combination 3 and 4 has initial sensitivity 70% & 75% respectively (at threshold 0.5) and initial FA 18 for both. However, even after threshold is increased to
. 9999 sensitivity is at 90% and FA becomes 85 and 113 respectively. This information can help us select appropriate combinations of tests within the budget. Similarly, for the application involving budget on total patient discomfort, we choose a total budget of 10, which is an upper bound on total discomfort values to be removed. Fig. 5 shows 15 options for se-lecting tests (blank entries in each column are selected for testing), and Fig. 6 shows the false abnormal obtained for each choice out of a total 738 patients. In our case, option 7 yields the minimum number of false abnormal. This technique can be used to provide patients with a choice to prioritize his/her own tests.
We observe after combining the result of both the applications that ‘Neck Nodes’, ‘Oral Hygene’ and ‘Reflux Gastritis’ are most prominent features among these 15 BCTs as they are common in almost all the options of 100% sensitivity. Also, it STA-21 is clear that the technique proposed here produces significant benefits over arbitrary test selection practiced today as false abnormal rates vary widely between different options of test selections. To the best of our knowledge, this is the first study of automatic diagnosis system which focuses on making false normal zero, and studying the consequent false abnormal rates.
5. Conclusion and future work
Research on prediction of diseases from EMR using machine learning techniques has been prevalent in the contemporary literature. This study illustrates that a machine learning based algorithm, kernel
logistic regression, can facilitate in the prediction of esophageal cancer relying on demographic, lifestyle, medical history and customized clinical test, with a very high accuracy upto 99.18% with a sensitivity nearing 100%. We also introduce the novel concept of selecting clinical tests, using preferences given by the patient, from the given set of tests. This brings a new dimension to all stakeholders in health industry (hospitals, patients, insurance providers) towards optimization of cost and freedom of selection without compromising the detection of all true patients.
In summary, we think that the outcome of this study is both methodological innovation and potential societal advancement: the former in terms of applying a new prediction technique for classifying esophageal cancer patients with tunable tradeoff between sensitivity and accuracy, and the latter allowing to choose the clinical tests as per wish of either patient, doctor or service provider.
The applicability of the machine learning techniques used in this study for other diseases and their appropriate features, is an interesting research direction with potential for making the healthcare system more effective. This work started with the objective of achieving 100% sensitivity with minimum False Abnormal for any test set after tuning with validation data set, which is yet to be achieved. We have reached to near zero FN but could not make inversion zero and hence a future work is recommended in this direction. Another potential research direction is to develop advanced sampling systems, which can potentially achieve sensitivity arbitrarily close to zero.
Asymmetric bagging and random subspace for SVM based relevance feedback , deep multimodal distance metric learning  or mul-tiview hessian LR method  may be applied tailoring the metho-dology to make these suitable for the stated dataset. In real life, samples are generally corrupted so a classifier which works optimum way with noisy level may be an appropriate one to enhance this research further. For example, importance re-weighting classification with noisy labels