Could Machine Learning Help HIV Treatment Programs? Three Questions to Consider


machine learning graphicsMachine learning is a ubiquitous and powerful tool used to solve problems across the medical field. However, despite its powerful capabilities and broad applicability, machine learning is not always the right solution to problems. So when is the right solution?

To answer this, we need to ask three key questions. Let’s start with a problem that countries frequently face in her HIV treatment programme: HIV treatment interruptions, where patients discontinue her HIV treatment plan for a variety of reasons. Here we are interested in understanding why patients discontinue treatment regimens. Who is likely to have their HIV treatment interrupted in the future? And can you predict what will happen and understand why it happens?

Let’s see if machine learning can help answer these questions.

Are you trying to understand relationships in your data, or are you trying to predict something about the future?

A clear problem statement is the foundation of an effective machine learning solution. If your goal is to identify relationships in your data to retrospectively understand who stopped HIV treatment and why, then descriptive and diagnostic analyzes fit your needs. Traditional statistical techniques such as regression analysis are well suited for examining associations between variables of interest. Simple descriptive statistics can also identify groups of patients who experience treatment interruptions more frequently than others.

However, if the problem statement is forward-looking, i.e., interested in identifying people most likely to experience HIV treatment interruptions and removing barriers to HIV treatment, machine learning A forecasting tool such as Because the problem is future-oriented, it is more suitable for predictive analysis than descriptive analysis. To build a machine learning solution to this problem, we need a historical patient record with an outcome variable (label) that captures whether the patient has experienced HIV treatment discontinuation. In this case, you are trying to predict a class of patients (those who are more likely to experience an interruption or less likely to experience an interruption), so it is important to have a sample of patient data where you know the outcome you are trying to predict. However, certain machine learning problems do not require outcome variables and are used to identify patterns within large datasets. For example, classify customers into different groups based on their buying habits. Such methods are called unsupervised machine learning because they have no outcome variable (label) to guide the model.

For example, Data.FI, one of the projects funded by PEPFAR and implemented by USAID, is developing a machine learning solution for Nigeria’s HIV testing service. The goal is to optimize HIV testing resources by using patient characteristics to predict who is most at risk of contracting HIV. The country has millions of patient records about who tested positive or negative, socio-demographic information, sexual behavior, marital status, and STD status, and machine learning models learn from this data. can be used to generate the probability of testing positive (HIV risk score). ) considers a range of patient characteristics. When a new patient comes in for an HIV test, healthcare professionals can enter the patient’s information into this model to generate her HIV risk score and decide who gets tested.

Is your data large and multidimensional?

Even if you’re trying to use data to predict something about the future, machine learning may not be right for you. The advantage of this tool is that you can learn complex interactions that you cannot identify on your own because they are too time consuming or the relationships are too complex for the human brain and statistical tests to understand and learn.

When we have a large number of entries in a dataset and we are interested in examining the relationships between many variables in each entry, machine learning unlocks many analytical opportunities not possible with descriptive analytics. For example, given 30-50 PEPFAR Monitoring, Evaluation, and Reporting (MER) metrics, existing data quality tools that use statistical analysis look at bi-directional ratios between metrics. What they won’t do is look at 30 different ratios.

Data.FI has developed an anomaly detection tool that can examine patterns across 50 variables at once. This is not humanly possible and cannot be done in Excel. Machine learning approaches like recommender systems can see patterns across all these variables and, based on those patterns, predict the most likely outcomes for variables with missing responses. However, if your dataset only has 5 He MER indicators, you probably don’t need machine learning.

On the other hand, if your dataset is small and you have only a few MER indicators, these statistical analysis tools can be very helpful in identifying historical patterns in your data. For this kind of use case, the extra cost and commitment to use machine learning probably isn’t worth it.

Do you have the people and information architecture to implement that work?

Launching and maintaining machine learning models requires staff with specialized skill sets and an information architecture that delivers machine learning-generated insights to decision makers when and how they need them. Is required. Specific skills and architecture depend on your use case. To build and evaluate the models, of course, you need someone with machine learning expertise. However, for a model to be useful, it must be deployed in some form. This format ranges from infrequently run analyzes available to a small group of central users to real-time data available to distributed networks of clinicians such as the Palladium-led models developed by his Data.FI and the Kenya Health Management Information System. decision-making support. A project (KeHMIS) directly integrated into the electronic medical record.

For the former, a simple server with a machine learning library installed is sufficient. In fact, with just a few clicks, we can rent space on cloud servers that provide what we need. For the latter, different techniques may be required. If internet connectivity is available, some models require an application programming interface (API) to connect to a centrally hosted model. Some models require you to install the model offline on your mobile device. Others may be familiar with PMML, ONNX, POJOs or MOJOs, how to transform a machine learning model so that it can communicate with Java-based applications, or others that may be familiar to some. You may need to use other fun acronyms that are barriers for people.


So is machine learning right for your project? It depends! This tool is great for learning from large amounts of labeled historical data to understand what is most likely to happen in the future. Machines can do things humans can’t, but machine learning isn’t always the answer or the complete answer to the problem. Moreover, even with a good problem statement and data, this approach may not be the right solution if the country does not have the prerequisite infrastructure and skilled manpower to maintain the machine learning solution.

If you’re already doing machine learning, you should evaluate it. The Data.FI Dashboard will show you how to do this.

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *