Binary classification

Binary classification is the task of classifying the members of a given set of objects into two groups on the basis of whether they have some property or not. Some typical binary classification tasks are

Statistical classification in general is one of the problems studied in computer science, in order to automatically learn classification systems; some methods suitable for learning binary classifiers include the decision trees, Bayesian networks, support vector machines, and neural networks.

Sometimes, classification tasks are trivial. Given 100 balls, some of them red and some blue, a human with normal color vision can easily separate them into red ones and blue ones. However, some tasks, like those in practical medicine, and those interesting from the computer science point-of-view, are far from trivial, and may produce faulty results if executed imprecisely.

Contents

Hypothesis testing

In traditional statistical hypothesis testing, the tester starts with a null hypothesis and an alternative hypothesis, performs an experiment, and then decides whether to reject the null hypothesis in favour of the alternative. Hypothesis testing is therefore a binary classification of the hypothesis under study.

A positive or statistically significant result is one which rejects the null hypothesis. Doing this when the null hypothesis is in fact true - a false positive - is a type I error; doing this when the null hypothesis is false results in a true positive. A negative or not statistically significant result is one which does not reject the null hypothesis. Doing this when the null hypothesis is in fact false - a false negative - is a type II error; doing this when the null hypothesis is true results in a true negative.

Evaluation of binary classifiers

See also: sensitivity and specificity

To measure the performance of a medical test, the concepts sensitivity and specificity are often used; these concepts are readily usable for the evaluation of any binary classifier. Say we test some people for the presence of a disease. Some of these people have the disease, and our test says they are positive. They are called true positives (TP). Some have the disease, but the test claims they don't. They are called false negatives (FN). Some don't have the disease, and the test says they don't - true negatives (TN). Finally, we might have healthy people who have a positive test result - false positives (FP). Thus, the number of true positives, false negatives, true negatives, and false positives add up to 100% of the set.

Specificity (TNR) is the proportion of people that tested negative (TN) of all the people that actually are negative (TN+FP). As with sensitivity, it can be looked at as the probability that the test result is negative given that the patient is not sick. With higher specificity, fewer healthy people are labeled as sick (or, in the factory case, the less money the factory loses by discarding good products instead of selling them).

Sensitivity (TPR) is the proportion of people that tested positive (TP) of all the people that actually are positive (TP+FN). It can be seen as the probability that the test is positive given that the patient is sick. With higher sensitivity, fewer actual cases of disease go undetected (or, in the case of the factory quality control, the fewer faulty products go to the market).

The relationship between sensitivity and specificity, as well as the performance of the classifier, can be visualized and studied using the ROC curve.

In theory, sensitivity and specificity are independent in the sense that it is possible to achieve 100% in both (such as in the red/blue ball example given above). In more practical, less contrived instances, however, there is usually a trade-off, such that they are inversely proportional to one another to some extent. This is because we rarely measure the actual thing we would like to classify; rather, we generally measure an indicator of the thing we would like to classify, referred to as a surrogate marker. The reason why 100% is achievable in the ball example is because one measures redness and blueness by detecting redness and blueness. However, indicators are sometimes compromised, such as when non-indicators mimic indicators or when indicators are time-dependent, only becoming evident after a certain lag time. The following example of a pregnancy test will make use of such an indicator.

Modern pregnancy tests do not use the pregnancy itself to determine pregnancy status; rather, they use human chorionic gonadotropin, or hCG, present in the urine of gravid females, as a surrogate marker to indicate that a woman is pregnant. Because hCG can also be produced by a tumor, the specificity of modern pregnancy tests cannot be 100% (in that false positives are possible). And because hCG is present in the urine in such small concentrations after fertilization and early embryogenesis, the sensitivity of modern pregnancy tests cannot be 100% (in that false negatives are possible).

In addition to sensitivity and specificity, the performance of a binary classification test can be measured with positive (PPV) and negative predictive values (NPV). The positive prediction value answers the question "If the test result is positive, how well does that predict an actual presence of disease?". It is calculated as (true positives) / (true positives + false positives); that is, it is the proportion of true positives out of all positive results. (The negative prediction value is the same, but for negatives, naturally.)

One should note, though, one important difference between the two concepts. That is, sensitivity and specificity are independent from the population in the sense that they don't change depending on what the proportion of positives and negatives tested are. Indeed, you can determine the sensitivity of the test by testing only positive cases. However, the prediction values are dependent on the population.

Example

As an example, say that there is a test for a disease with 99% sensitivity and 99% specificity. Say 2000 people are tested, and 1000 of them are sick and 1000 of them are healthy. One is likely to get about 990 true positives, 990 true negatives, and 10 of false positives and negatives each. The positive and negative prediction values would be 99%, so the people can be quite confident about the result.

Say, however, that of the 2000 people only 100 are really sick. Now you are likely to get 99 true positives, 1 false negative, 1881 true negatives and 19 false positives. Of the 19+99 people tested positive, only 99 really have the disease - that means, intuitively, that given that your test result is positive, there's only 84% chance that you really have the disease. On the other hand, given that your test result is negative, you can really be reassured: there's only 1 chance in 1882, or 0.05% probability, that you have the disease despite your test result.

Measuring a classifier with sensitivity and specificity

Suppose you are training your own classifier, and you wish to measure its performance using the well-accepted metrics of sensitivity and specificity. It may be instructive to compare your classifier to a random classifier that flips a coin based on the prevalence of a disease. Suppose that the probability a person has the disease is p and the probability that they do not is q=1-p. Suppose then that we have a random classifier that guesses that you have the disease with that same probability p and guesses you do not with the same probability q.

The probability of a true positive is the probability that you have the disease and the random classifier guesses that you do, or p^2. With similar reasoning, the probability of a false negative is pq. From the definitions above, the sensitivity of this classifier is p^2/(p^2%2Bpq)=p. With more similar reasoning, we can calculate the specificity as q^2/(q^2%2Bpq)=q.

So, while the measure itself is independent of disease prevalence, the performance of this random classifier depends on disease prevalence. Your classifier may have performance that is like this random classifier, but with a better-weighted coin (higher sensitivity and specificity). So, these measures may be influenced by disease prevalence. An alternative measure of performance is the Matthews correlation coefficient, for which any random classifier will get an average score of 0.

Converting continuous values to binary

Tests whose results are of continuous values, such as most blood values, can artificially be made binary by defining a cutoff value, with test results being designated as positive or negative depending on whether the resultant value is higher or lower than the cutoff.

However, such conversion causes a loss of information, as the resultant binary classification does not tell how much above or below the cutoff a value is. As a result, when converting a continuous value that is close to the cutoff to a binary one, the resultant positive or negative predictive value is generally higher than the predictive value given directly from the continuous value. In such cases, the designation of the test of being either positive or negative gives the appearance of an inappropriately high certainty, while the value is in fact in an interval of uncertainty. For example, with the urine concentration of hCG as a continuous value, a urine pregnancy test that measured 52 mIU/ml of hCG may show as "positive" with 50 mIU/ml as cutoff, but is in fact in an interval of uncertainty, which may be apparent only by knowing the original continuous value. On the other hand, a test result very far from the cutoff generally has a resultant positive or negative predictive value that is lower than the predictive value given from the continuous value. For example, a urine hCG value of 200,000 mIU/ml confers a very high probability of pregnancy, but conversion to binary values results in that it shows just as "positive" as the one of 52 mIU/ml.

See also

Bibliography