Medical treatments are prescribed to alleviate illness, but sometimes their effects can be harmful. A prescribed treatment can lead to an unexpected adverse drug reaction (ADR), causing fresh or further harm. This is typically the result of the incorrect medication or dosage being prescribed, or it can occur because of an interaction of a combination of drugs. ADRs differ from side effects, which can occur even when the correct medication has been prescribed, and are usually expected, mild and self-resolving (although some side effects can be severe).

Deborah Weinswig

Incorrect prescriptions may sometimes be the result of misinformed or underinformed decision making. Health care professionals tend to be educated on potentially harmful drug combinations and instances that can cause ADRs, but memorizing the ever-growing volume of information and keeping up to date on the uses and effects of new drugs can prove tedious and challenging. Moreover, because of the large number of patients that health care providers see every day, it is possible that they may fail to identify the likelihood of such an ADR.

People can turn to public resources such as the Drugs.com website, which allow them to do a drug interaction check, whereby they enter the name of a drug and check for possible interactions with combinations of other drugs, food and beverages, and diseases or conditions. However, the information provided is not uniquely tailored to individuals.

This information gap is where artificial intelligence (AI) and machine learning (ML) come in, helping health care providers to better predict — and, so, better prevent — the occurrence of ADRs at the individual patient level.

Causes of ADRs

While some drug combinations are known to cause adverse reactions, in many circumstances it is difficult for health care providers to predict such reactions, as many variables are at play. These include:

• Physiological factors — A patient’s age, gender, body weight and composition, as well as a history of allergies, can affect how a drug behaves in the body.

• Presence of multiple health conditions — Some people may have multiple health conditions, of which one could inhibit the efficacy of a drug meant to treat another condition.

• Lifestyle and social factors — Individual diets, lifestyles and social factors can adversely affect the absorption of prescribed drugs.

• Negligence — Patients may fail to take medicines in the prescribed doses, while health care providers may fail to take necessary precautions or study patients’ history adequately. They may also offer patients improper advice for taking prescribed drugs or fail to administer the drugs properly.

• Patients’ failure to provide complete information — Some patients may inadvertently or intentionally fail to provide complete information about their health history and lifestyle habits to health care providers.

• Drug interactions — Some drugs can positively or negatively affect the functioning of another drug. If a drug inhibits the functioning of another drug, it can result in an ADR.

Prevalence of ADRs in U.S.

Some of the signs of ADRs, such as skin rashes and dizziness, can be severe on some occasions, and they can result in life-threatening conditions. Every year, more than 1 million people are admitted to U.S. hospital emergency departments for ADRs, according to a research paper published in the Journal of the American Medical Association in November 2016. More than one-quarter of these patients need to be hospitalized for further treatment, the research found.

ADRs result in significant costs in the U.S., where an estimated $200 billion is spent annually on unnecessary or improperly prescribed drugs and related medication, according to health care market research firm IQVIA. To put this into context, the Centers for Disease Control and Prevention reported that the U.S. spent $3.3 trillion on health care in 2016.

The scale of drug prescriptions contributes to the prevalence of ADRs. More than half of Americans regularly take prescription drugs — four on average — and 75% of them take at least one over-the-counter drug regularly, according to an April 2017 Consumer Reports survey of 1,947 American adults. Some 53% of prescription-drug users reported that they received their prescriptions from more than one provider, and 35% stated that a health care provider had never reviewed their medication to see if they could stop taking it.

ML can reduce likelihood of ADRs

AI and ML are human-like or intelligent behaviors exhibited by machines. AI is enabled by computer programs that help machines make independent decisions through cognitive functions similar to those typically exhibited by human beings, such as learning, decision making and problem solving. Within the area of AI, ML refers to the techniques used to perform cognitive functions.

AI and ML programs or algorithms can easily evaluate vast, unstructured data sets; study multiple variables; identify patterns; and calculate the probability of an event occurring and its outcome. These technologies are built into computer software that is designed to assist health care providers with decision making, such as clinical decision support systems.

Various ML algorithms have been created for use across different industries and sectors. These are divided into three broad classifications:

• Supervised ML algorithms — These programs make predictions based on a given set of samples, where the outcome is known or clearly defined. For example, a machine may be asked to pick out red blocks and blue balls from a collection of objects in different colors.

• Unsupervised ML algorithms — When there are no predefined labels, these algorithms process and sort data into clusters based on similarity between the data and the frequency of occurrence of associated items. For example, a program may assess a set of photographs of people and identify those who are smiling and those who are not based on studying facial patterns. In the case of a shopper who buys a bag of coffee online, a website’s built-in program may present the shopper with associated products such as milk and sugar, because of an algorithm that has “learned” to do so based on similar purchase patterns from the same customer or other customers in the past.

• Reinforcement ML algorithms — These algorithms make decisions based on constant feedback and the effectiveness of the decision made, and then devise strategies to provide better outcomes. For example, they may be used to understand the best moves in a game of computer chess or to learn the best routes and speeds for self-driving cars.

There are many ML algorithms built for specific sectors, uses and types of data under these broad classifications. An unsupervised ML algorithm called Apriori is currently the most suitable for identifying potential ADRs, given the nature of patient data, the calculations that need to be made and the desired outcome.

How Apriori works

The Apriori algorithm works by creating association rules for a given data set that are generated in an “if, then” format — i.e., if A happens or is present, then B must happen or be present with a certain probability; or if A and B occur frequently, then C must occur frequently with a certain probability, and so on. The Apriori algorithm works on two basic principles:

• All subsets of an item set that occurs frequently also occur ­frequently.

• If an item set occurs infrequently, its supersets also occur infrequently.

Let us consider the example of warfarin (sold under the brand names Coumadin and Jantoven), a commonly prescribed drug to thin blood and treat blood clots or lower the chance of heart attack or stroke in some people.

Some of the side effects of warfarin include dizziness, excessive bleeding upon injury and severe headaches. People of Asian descent or age 60 or older could also experience more side effects that could potentially be fatal.

In terms of ADRs, warfarin is known to significantly interact with 214 drugs and drug combinations, as documented by online drug interaction checker Drugs.com. Consuming food rich in vitamin K, a nutrient that promotes clotting, could reduce the efficacy of warfarin. As a blood thinner, warfarin could adversely affect patients with severe diabetes or hypertension, as they may be at increased risk for hemorrhage.

Apriori algorithm application

When prescribing warfarin, a doctor needs to consider various factors about a patient in order to determine the correct dosage. Two patients may have identical physiological features but have different ethnicities which could affect the doctor’s decision. It is also possible that one patient may already be taking other medications and supplements that the second patient is not, or that the two patients lead very different lifestyles. These factors, along with many others specific to each patient, form the item sets and subsets that the physician must consider before prescribing a drug.

A clinical decision support system that uses the Apriori algorithm can run through each patient’s medical history and consider all combinations of their individual characteristics, the medications they are taking and other variables. The algorithm helps the clinical decision support system calculate the probability of a particular ADR occurring and thereby helps the health care provider decide the appropriate prescription and treatment for each patient.

Of course, not every combination of drugs is harmful. However, the more drugs a person takes, the higher the chance of an ADR occurring, and studies have shown that a significantly large number of Americans take more than one drug regularly.

Limitations of ML
in identifying ADRs

Existing challenges of ML applications to help reduce the occurrence of ADRs include:

• Inadequate sample sizes — Data migration from paper to electronic form has certainly helped create vast data sets, but these may be limited to commonly occurring conditions. Rarer conditions constitute a relatively smaller data set, resulting in a sample size of variables that may be inadequate to process wide-ranging combinations and generate sufficient probabilities to assist health care providers effectively.

• The need for human intervention — ML algorithms can only assist health care providers in making a decision; the algorithms cannot decide the appropriate medication and dosage for a patient on their own.

• Biased data — Because ML algorithms produce probabilities based on associations between variables in a data pool, the data should be unbiased. Providing a system with excess variables for some drugs or conditions, or with too few variables, can result in manipulated probabilities of certain ADRs ­occurring.

Health care providers constantly need to make critical decisions, and they may not always be able to recollect all the information that they need. Learning about new drugs and complicated conditions is also time consuming, and many doctors and pharmacists already find themselves under tight time constraints.

A handy repository of relevant information is helpful for health care providers, especially when they are treating numerous patients or those with multiple or serious health conditions. Programs that can crunch unique variables and provide the probability of a treatment’s efficacy or negative impact on a patient are beneficial to health care professionals who are making crucial decisions. However, given the current limitations of ML tools and the possibility of them generating biased results, we think that the technology has not yet advanced far enough to function independently. Until that happens, health care providers will need to continue to use ML algorithms as secondary decision-making tools to support their own insight and expertise.

Deborah Weinswig is founder and chief executive officer of Coresight Research. She can be contacted at [email protected].