Based on "Module 3, Section 2: AI Modeling, Machine Learning on Steroids" from the course Demystifying AI for Health Professionals, presented by Vahab Khademi, PhD, Postdoctoral Researcher at The George Washington University.

As a healthcare professional, you are increasingly asked to trust, and sometimes act on, the output of machine learning models. But there is a common belief that the models themselves belong to people who write code for a living. In Part Two of our hands-on machine learning module, we set out to challenge that belief using a free, visual tool called Orange Data Mining. With Orange, you build a model not by typing commands, but by dragging components onto a canvas and connecting them, like wiring together a flowchart that happens to run. In this walkthrough, written for clinicians rather than programmers, we will install the software, learn how it is organized, and build a model that predicts body fat from patient measurements. But before we drag a single widget, it is worth understanding where this quirky little tool came from and why it belongs in your hands.

Where Orange Came From, and Why It Is Free

Orange is not a commercial product with a marketing budget. It is an open-source project that has been developed at the University of Ljubljana in Slovenia since 1996, which makes it nearly as old as the modern field of data mining itself. It remains free to download and use, for one simple reason: it was built to teach.

The idea behind it is older still. It traces back to Donald Michie, a British pioneer of artificial intelligence who, in the late 1990s, argued that machine learning needed an open, shared toolbox so that anyone, not just specialists, could experiment with data. That conviction is baked into the software's whole personality. The team that built it even titled their own history of the project "Data Mining Fruitful and Fun," which tells you something about their intentions. It also explains the icon you are about to meet: a cheerful cartoon orange wearing thick black glasses. The playfulness is the point. The goal was never to intimidate; it was to invite people in.

Why This Belongs in a Clinician's Toolkit

Orange was designed for exactly the situation many clinicians find themselves in: you have real questions and real patient data, but you do not have the time or the inclination to become a programmer. Its visual, drag-and-drop approach lets you carry out serious analysis, training models, comparing them, and visualizing the results, without writing a line of code. It is no coincidence that the tool has found a home in biomedicine, bioinformatics, and genetics, and above all in teaching, where it lets people see the machinery of machine learning instead of simply trusting it.

That is the real reason it matters to you. You do not need to build AI to be affected by it; these tools are already entering imaging, risk prediction, and clinical decision support. But understanding how a model is assembled, what goes in, what comes out, and where it can quietly go wrong, is exactly what separates a passive user from a clinician who can evaluate a tool before trusting it with a patient. Orange is a low-stakes, hands-on place to build that judgment, using the kind of health data you already understand.

Getting the Software: Installing Orange Data Mining

Orange is available for both Windows and Apple machines. To get it, open any web browser and search for "orange data mining." The first result takes you to the project's website, where you click "Download Orange."

From there, you will see two options, one for Windows and one for macOS, and each has a sub-option. On Windows, you can either download and install the software, or download a portable version that runs without installation. On macOS, your choice depends on your processor. Newer Mac computers use Apple's own ARM 64 architecture, so you would choose "Orange for Apple silicon," while older machines use "Orange for Intel." Before downloading the Mac version, make sure you know which CPU your computer uses. In this walkthrough we are on a Windows machine, so we download Orange for Windows, locate the downloaded file, and install it.

Once the software is installed, you can double-click the file to open it. And there it is: that smiling orange with the thick black glasses.

Finding Your Way Around the Workspace

The first thing you see is a welcome screen. From here you can start a new project, open an existing one, or jump back into your most recent work. If you are completely new to the tool, this screen also points you toward tutorials on YouTube, documentation, help files, and examples to get you started. Once you close this splash screen, you arrive at the main workspace.

The interface has two main parts. The largest area is the canvas. Think of it like a drawing canvas, except instead of drawing, you place machine learning components onto it and connect them together. Those components are called widgets, and you will find them in the panel on the left.

Here is the clever part, and the reason Orange is such a good teacher: the widgets are not thrown together randomly. They are arranged to follow the natural workflow of a machine learning project, which always begins with a question and some data. As you move down the panel, you are essentially reading the life cycle of an analysis from start to finish:

• Data: Widgets for bringing in your data, whether from a file already on your computer, a URL, an imported CSV, or a dataset built into the software.
• Transform: Tools for reshaping your data, such as selecting only certain columns or rows, or merging two datasets together.
• Visualize: Plots for exploring your data, including scatter plots, line plots, and bar plots.
• Model: The heart of the toolkit, containing supervised models such as random forests, gradient boosting, linear regression, and logistic regression.
• Evaluate: Tools for checking how well a model performs, such as Test and Score, ROC analysis, and the confusion matrix.
• Unsupervised: Models that work with unlabeled data so the machine can find patterns on its own, such as clusters, groups, relationships, or latent dimensions.

Beyond these, there are also specialized widgets for image analytics, bioinformatics such as genetics, educational applications, and text work like text prediction or sentiment analysis. The takeaway is that the layout itself is a lesson: by the time you have moved a project across the canvas, you have walked the entire arc of a real machine learning workflow.

Starting With a Question and Some Data

The first method we will demonstrate is linear regression, and our goal is to predict body fat. Following the machine learning workflow, we start by getting the data. For this example, we use a built-in dataset from Kaggle that comes packaged with the software, which spares us the hunt for a file to practice on.

We drag the Datasets widget onto the canvas, double-click it, and search for "body fat." The search results show us a dataset of 252 patients (called instances), with 16 variables. The target is numeric, meaning we are predicting a number, and the dataset belongs to the health sciences. A short description appears below it. To download the data onto your machine, double-click the name of the dataset. Once it has downloaded, the circle next to it turns green. If it is not green, the data is not yet on your computer, so double-click and wait a moment for it to finish.

A good habit at this stage is to rename your widgets so the canvas stays readable. Since our dataset is about body fat, we right-click the widget, choose rename, and call it "body fat data." It is a small thing, but as your workflows grow, clearly labeled pieces are what keep them from becoming a tangle.

Looking at the Data Before You Model It

Before building anything, it is worth seeing what is actually inside the dataset, because a model is only ever as trustworthy as the data feeding it. We drag a connection out from the body fat widget and choose Data Table from the list that appears. (You can also drag the Data Table widget directly onto the canvas, but it will start empty until you connect the data to it.) Double-clicking the Data Table populates it, and now we can see the variables: body fat, patient ID, age, weight, height, and other measurements. These are the features, or predictors, that we will use to predict body fat.

You may notice that the patient ID column is shown in a different color. That is because it is not used in the analysis; it is treated as metadata. This is a meaningful distinction, not a cosmetic one. A patient's ID number carries no predictive information, and you would not want a model trying to learn from it. The information panel on the left confirms the shape of the data: 252 instances with no missing values, 14 features, a numeric outcome, and one meta attribute, the patient ID, which will sit out of the analysis.

To make the roles of each variable explicit, we can go to Transform and use Select Columns, connecting it to a Data Table. Opening it shows clearly which variables are features and which one is the target or outcome, and it confirms that the patient ID is a meta variable, set aside from the analysis.

Training the Linear Regression Model

Now we are ready to build the model. We have our data, we know the target variable, and we know our features. The choice of linear regression is not arbitrary: because the target, body fat, is a continuous number rather than a category, a regression model is the right family for the job. Matching the model to the kind of question you are asking is one of the most important habits in machine learning, and Orange makes that choice visible.

From the Model group, we search for Linear Regression and drag it onto the canvas. We then connect the Select Columns widget to it. This step, where the model learns from the data, is called fitting or training the model. Double-clicking the widget reveals the technical parameters of the model. For this module we leave these alone; they are there for more advanced users who want to experiment, and one of the nice things about Orange is that it lets beginners ignore that complexity until they are ready for it.

Evaluating How Well the Model Performs

A trained model is only useful if it can make good predictions on data it has not seen before, so the final step is evaluation. This is the part people are most tempted to skip, and the part that matters most. From the Evaluate group, we use the Test and Score widget and drag it onto the canvas. It needs to know two things: which model to test, so we connect our Linear Regression to it, and what data to test it against, so we supply data for it to evaluate the trained model on.

This is the moment the whole canvas comes alive. The data flows in, the model learns from it, and the evaluation tells us whether its predictions can actually be trusted, all laid out visually so you can follow the logic from one end to the other.

The Turning Point: When the Black Box Becomes Glass

It is easy to imagine machine learning as something sealed inside code that only specialists can read. Orange was built, quite deliberately, to make the opposite point. Every stage we walked through, getting the data, inspecting it, choosing a model that fits the question, training it, and then testing how well it performs, is a decision you can see, arrange, and reason about. That is no accident; it is the founding idea Donald Michie sketched out decades ago, finally made practical. When you can watch the data move from one widget to the next, the model stops being a mysterious black box and becomes something closer to glass. And that clarity is exactly what lets you ask sharper questions about the AI tools arriving in your own practice, and judge for yourself whether their answers make sense for the patient in front of you.

About the CAIT Center

The Collaborative AI Technology Center (CAIT Center) is a research-based partnership between the GW Biomedical Informatics Center and the University of Maryland Eastern Shore School of Pharmacy. We provide the educational tools clinicians need to understand, evaluate, and responsibly apply AI in healthcare.

Keep Building Your Skills

This hands-on walkthrough is part of our broader effort to make machine learning approachable for professionals, no coding required.

Course: Demystifying AI for Health Professionals

• Hands-On Practice: Build and evaluate real machine learning models using free, visual tools like Orange Data Mining.
• No Code Required: Focus on the workflow and the meaning of your results, not on programming.
• Credly Digital Badge: Earn a formal certification to show you can work confidently with these tools.