Medications are another type of commonly used data in Computational Phenotyping Algorithms. Data like diagnosis codes and laboratory tests, may only be received during the initial diagnosis of a disease. By looking at the treatments for a disease, you may identify patients who are diagnosed many years before receiving care in your health system. The first challenge with using medications for Computational Phenotyping, is that medications may be used to treat many different disorders. The reason a patient is taking a drug is called the "Indication". Some drugs like insulin only have a single indication, diabetes. Others may only have a single approved use, but are used off-label to treat other conditions. For example, spironolactone is approved for the treatment of congestive heart failure. However, it also has potent anti-angiogenic effects, which can help women with polycystic ovarian syndrome, control some of the negative side effects of the disease. In some cases, these off-label uses of the drug are actually more common than the original approved indication. Because of these multiple indications, it's recommended to use medications only in combination with other data types. Using the presence of a medication in your phenotyping algorithm is relatively easy. Most prescriptions are available as structured data. Things get more complex if your algorithm requires specific dosages of a therapy. Well, this may be feasible for drug therapies where the dose is in a structured form in the original prescription. Some drugs actually have very complicated dosing schedules that are simply not available at the time of prescription. For example, warfarin, which is a blood thinner, is extremely complex. The dose changes over time. Warfarin is empirically dosed, which means that patients are started on an initial dose of the drug, and then providers measure how well their blood clots and adjust their dose. This process of testing and adjustment occurs until the right effect is reached. Getting to the right dose may be difficult, and patients may have to take different doses each day of the week. This complexity is demonstrated by the fact that there are nine different pill dosages, and the pills have to be color-coded to make sure the patients take the right pill on the right day. Warfarin doses are unique, in that their doses are typically stored in notes or clinic forms written by the anticoagulation clinic. Finding and interpreting these dosing instructions is complex, and may require natural language processing. Another complication of using medication data for Computational Phenotyping, is that it is difficult to tell when a drug therapy has been discontinued. If you were simply trying to determine if the patient has ever had a disease, this may be acceptable. However, if you are trying to identify drug side effects, additional effort will be required to see if the side effect occurred while the patient was still on the drug. Overall, medications are a powerful tool for Computational Phenotyping Algorithms. But as with other data types, they have limitations and make them best used in combination with other data elements.
