One moment creates almost all of them
A patient arrives. The receptionist searches, does not find them (spelling, a different phone number, a name entered as initials last time) and creates a new record. Four seconds, no error message, and the patient now has two histories.
Every subsequent problem descends from that moment: the doctor sees half a history, the follow-up attaches to the wrong case, the outstanding balance is on the other record, and the analysis at the end of the year counts one person as two.
Indian names make search harder, not impossible
Transliteration varies, initials expand differently, the same name appears with and without a surname, and a family often shares one mobile number. A search that expects exact spelling will fail routinely in this environment.
The practical answer is not a cleverer search box. It is to search on the things that vary least, the phone number and the identifier, and to surface near-matches at the moment the receptionist is about to create a second record.
Detection has to interrupt, gently
Merging duplicates later is a clean-up project that hospitals put off indefinitely, because it is tedious and because merging clinical records carries its own risk. Preventing them costs a two-second glance.
The interruption has to be at the right strength. A hard block turns a legitimate new registration into a support call. A warning nobody reads may as well not exist. Showing the possible matches, with name, phone and last visit, and letting a human decide, is the version that works, because the receptionist usually recognizes the patient the moment she sees the last visit date.
What it costs to leave alone
Clinically, a doctor working from half a history is the real cost, and it is invisible until it is not. Commercially, duplicates hide outstanding balances, break follow-up recall, and make every count you report slightly wrong in a direction nobody can quantify.
The uncomfortable part is that a hospital with a duplicate problem usually does not know its size. The only way to find out is to look for near-matches across the whole register, and most systems will not tell you.