Using linear regression for classification problems is never a good choice.
Let me try to explain this by a very common example.
Lets say you have to predict whether the tumor is malignant or not based on the tumor size. Here the green line shown is our regression line.
So, in the above case, we can say that tumor size greater than 0.5 are malignant and rest are not. So if we easily done classification problem with linear regression.
But consider another case.
Here, if we classify that tumor size greater than 0.5 are malignant, that will not work here. We need to change the threshold, like 0.2 or something to make our predictions correct.
But we cannot change the threshold each time when a new sample arrives. Instead, our algorithm should learn it off from the training set data, and then make correct predictions for the data we haven't seen before.
Hope this will clear your doubt.
Further you can also refer to this discussion https://discuss.analyticsvidhya.com/t/using-linear-regression-for-a-classification-problem/9848