Learn to Build Basic Machine Learning Models from Health Data: A Hands-On Guide for Non-Coders

Abstract

The increasing integration of artificial intelligence in health care has created new opportunities for improving diagnosis, prognosis, and clinical decision-making. However, the adoption of machine learning (ML) by health care professionals remains limited due to the technical complexity associated with coding and data analysis. This review article aims to bridge this gap by providing a practical, hands-on guide for non-coders to build basic ML models using structured health data.

We describe a simplified four-step workflow consisting of data cleaning, model building and evaluation, external validation, and deployment for clinical use. To operationalize this process, we introduce four user-friendly, no-code applications developed using R and Shiny—CleanSight, MLSight, ValidateSight, and PredictSight. These tools enable users to preprocess data, train and evaluate ML models, and generate real-time predictions through intuitive graphical interfaces without requiring programming skills. This workflow is designed for structured/tabular clinical data and does not include computer vision tasks such as image classification or segmentation.

A case study using a fictitious dataset on nonalcoholic fatty liver disease is presented to demonstrate the complete workflow, including handling missing data, training a prediction model, and applying it in a simulated clinical scenario. The applications are designed to run on standard personal computers, making them accessible in routine health care and academic settings.

By simplifying complex ML processes and emphasizing practical usability, this guide empowers clinicians, including radiologists, to transform routine clinical data into actionable predictive tools. The approach has the potential to enhance data-driven clinical practice and promote wider adoption of ML in health care, particularly among users with limited technical expertise.

Introduction

The growing integration of artificial intelligence (AI) into health care is transforming clinical practice by enhancing diagnosis and decision-making. ¹ A key component of this transformation is machine learning (ML), which enables the identification of complex patterns within large datasets and supports predictive modeling based on prior clinical information. ² With the increasing digitization of health records and widespread use of electronic data systems, ³ clinicians now have access to substantial volumes of patient data that hold significant potential for generating actionable insights. ⁴

Despite these advancements, the practical application of ML in routine clinical settings remains limited. One of the primary challenges is the technical barrier associated with model development, which typically requires proficiency in programming languages, statistical concepts, and data handling techniques. ⁵ For many health care professionals, this creates a disconnect between data availability and data utilization. ⁶ As a result, valuable clinical datasets often remain underused, and opportunities for developing context-specific predictive tools are missed. ⁷

In response to this gap, there has been increasing interest in no-code and low-code solutions that simplify the ML process. These approaches aim to make model development more accessible by replacing complex coding requirements with intuitive graphical user interfaces. ⁸ By abstracting underlying computational processes, such tools allow users to focus on clinical interpretation rather than technical implementation, thereby facilitating broader adoption among nontechnical users. ⁹

With this background, this article aims to provide a practical, step-by-step guide for clinicians to understand and perform basic ML using health data through four no-code graphical applications. Although AI in radiology is often associated with image analysis, many clinically relevant applications use structured or tabular data, such as predicting imaging outcomes, contrast-related adverse events, and patient prognosis. This guide focuses on structured-data ML and excludes deep learning–based image analysis.

Basic Steps of Building a Machine Learning–Based Prediction Model

Before starting the ML process, it is essential to clearly define the clinical problem and ensure proper data entry. The objective of the model, such as predicting a disease, outcome, or risk, should be well specified. At the same time, relevant patient data must be collected in a proper format, typically in spreadsheets. ¹⁰ Careful and consistent data entry is crucial, as errors at this stage (e.g., missing values, incorrect units, or inconsistent formats) can directly affect model performance. ¹¹ Once the problem is defined and the dataset is properly organized, the ML workflow can be approached in four key steps.

Step 1: Splitting and Data Cleaning

The first step involves preparing the dataset for analysis. Clinical data often contain missing values, incorrect entries, or inconsistencies that need to be addressed before model building. ¹² Data cleaning ensures that the dataset is complete, accurate, and standardized. A clean dataset forms the foundation for any reliable ML model. ¹³ Some of the common data-cleaning methods are described in Table 1 .

Table 1. Some of the most common data cleaning methods.

Method	Meaning	Example
Remove duplicate records	Delete repeated entries of the same patient or observation	The same patient entered twice
Correct invalid values	Fix impossible values	Age = − 5 years corrected or flagged
Standardize data formats	Make data consistent	Dates converted to DD-MM-YYYY format
Remove unnecessary variables	Delete irrelevant columns	Registration number removed from analysis
Identify and review outliers	Check unusually high or low values	Height = 350 cm reviewed for error
Handle missing values	Fill in or remove missing information	Missing BMI value replaced with median BMI

Abbreviation: BMI, body mass index.

Among these, missing data are common in clinical datasets due to incomplete records, nonresponse, or data entry errors. If unaddressed, they can cause model training failures, introduce bias (especially when missingness is not random), and reduce sample size when cases are deleted. Therefore, proper handling is essential to ensure accurate and reliable ML models. Common approaches include removing records when missingness is minimal and random, or imputing values using simple methods such as mean, median, or mode. ¹⁴ Mean imputation may be appropriate for approximately normally distributed variables, whereas median imputation is generally preferred for skewed clinical data because it is less influenced by extreme values. Mode imputation may be used for categorical variables. ¹⁵

Users may first divide the dataset into training and testing subsets by specifying the desired proportion (e.g., 70:30 or 80:20). ¹⁶ This step is performed before missing-value imputation to minimize information leakage between datasets. Imputation parameters are learned from the training dataset and subsequently applied to the testing dataset using the same rules.

Step 2: Model Building and Evaluation

In this step, the cleaned dataset is used to train a ML model. The data are usually divided into two parts: a training set to teach the model and a testing set to evaluate its performance (commonly, 80% of the data are used for training and 20% for testing). ¹⁶ The model learns patterns from the training data and is then evaluated using the testing data, which the model has not seen before. This approach provides a more realistic estimate of how the model is likely to perform when applied to new patients in clinical practice.

To further improve the reliability of the model, the training data can be repeatedly divided into smaller groups, allowing the model to be trained and tested several times before the final model is developed. This process, known as cross-validation, helps ensure that the results are not dependent on a single data split and reduces the chance that the model merely memorizes the training data instead of learning meaningful patterns (overfitting). ¹⁷ In addition, many ML algorithms require users to specify certain settings before training. These settings, called hyperparameters, can be automatically tested in different combinations to identify those that provide the best model performance, a process known as hyperparameter tuning. ¹⁸

For models that predict the probability of an outcome (e.g., the probability of having a disease), an additional step called probability calibration can be performed. Calibration adjusts the predicted probabilities so that they more closely reflect the true likelihood of the outcome. ¹⁹ For example, among patients assigned a predicted risk of 80%, approximately 80% should actually experience the outcome if the model is well calibrated.

Common ML models in health care are listed in Table 2 . ^{20 21} Performance is measured using metrics such as accuracy, sensitivity, and specificity, helping determine how well the model performs in a clinical context ( Table 3 ). ^{22 23}

Table 2. Common machine learning models used in health care.

Model	Type	Typical use	Key advantage	Limitation
Linear regression	Regression	Predict continuous outcomes (e.g., blood pressure)	Simple and interpretable	Assumes a linear relationship between predictors and outcome
Logistic regression	Classification	Predict binary outcomes (e.g., disease vs. no disease)	Clinically interpretable and computationally efficient	Limited in capturing complex nonlinear relationships
Decision tree	Classification or regression	Risk stratification and prediction	Intuitive, rule-based, and easy to visualize	Prone to overfitting
Random forest	Classification or regression	General-purpose prediction	High accuracy and robust to overfitting; captures nonlinear relationships	Less interpretable than a single decision tree
Support vector machine	Classification or regression	High-dimensional data	Effective in high-dimensional feature spaces	Computationally intensive for large datasets; difficult to interpret
k-Nearest neighbors	Classification or regression	Small- to moderate-sized datasets	Simple and nonparametric	Sensitive to noise and feature scaling; computationally expensive for large datasets
Naïve Bayes	Classification	Text classification and probabilistic prediction	Fast, efficient, and performs well with high-dimensional data	Assumes conditional independence among predictors
Gradient boosting machine (GBM)	Classification or regression	Structured/tabular data	High predictive performance and flexible	Prone to overfitting if not properly tuned; slower training
Extreme gradient boosting (XGBoost)	Classification or regression	High-performance prediction on structured data	Excellent predictive accuracy, regularization, and handles missing values	Computationally intensive and requires careful hyperparameter tuning

Note: This table summarizes commonly used machine learning algorithms in health care and is not intended to be an exhaustive list.

Table 3. Common model-performance metrics.

Task category	Metric name	What it measures
Regression (predicting a continuous number, like blood pressure or age)	MSE (mean squared error)	The average squared difference between the predicted and actual numbers.
	RMSE (root mean squared error)	The square root of the mean of the squared differences between predicted and observed values, expressed in the same units as the outcome.
	MAE (mean absolute error)	The absolute average difference between predictions and actuals, without heavily penalizing large outliers.
	R-squared (R 2 )	The percentage of the outcome's variation that the model successfully explains (closer to 1.0 is better).
Classification (predicting a distinct category, like disease present/absent)	Accuracy	The overall percentage of predictions (both positive and negative) that the model got exactly right.
	Precision	Out of all cases the model called positive, the percentage that was actually positive.
	Sensitivity (recall)	Out of all actual positive cases, the percentage the model successfully found (critical for medical screening).
	Specificity	Out of all actual negative cases, the percentage the model correctly identified as negative.
	F1 score	A balanced score combining precision and sensitivity, highly useful if the condition you are predicting is rare.
	AUC-ROC	The model's overall ability to distinguish between the categories (a score of 0.5 is guessing, 1.0 is perfect).

Abbreviations: AUC-ROC: Area under the receiver operating characteristic curve; F1 score: harmonic mean of precision and recall.

Step 3: External validation

External validation is the process of evaluating a ML model using an independent dataset that was not used during model development or internal validation. ²⁴ This dataset is typically collected from a different hospital, geographical region, time period, or patient population. External validation assesses whether the model maintains its performance in new, real-world settings and determines its generalizability. A model that performs well only on the development dataset may not work reliably in other populations due to differences in patient characteristics, disease prevalence, or clinical practice. ²⁵ Therefore, external validation is considered an essential step before a ML model is adopted for routine clinical use.

Step 4: Deployment and Clinical Application

Once the model demonstrates satisfactory performance, it can be deployed for practical use. In this stage, the model functions as a decision-support tool, where new patient data are entered to generate predictions or risk estimates. This allows clinicians to apply insights derived from past data to current clinical scenarios. ²⁶ When implemented effectively, such models can enhance decision-making and support data–driven clinical practice.

A Toolkit for Basic Machine Learning

For building basic prediction models using structured clinical data (without images or high-dimensional inputs), a standard home computer is generally sufficient. The programming language R provides a powerful, free, and widely used platform for developing such models. ^{27 28} However, the requirement of coding can be a barrier for many clinicians. To address this, we developed a set of user-friendly applications using Shiny. ^{29 30} Through this interface, clinicians can clean their data using CleanSight, ³¹ build and evaluate ML models with MLSight, ³² and save the trained model. The saved model can then be validated externally by ValidateSight, ³³ and uploaded into PredictSight for real-time clinical use, enabling practical deployment without any coding requirement ( Table 4 ). ³⁴

Table 4. No-code toolkit and other tools for machine learning workflow.

Name	Brief description	URL
CleanSight	A tool for data cleaning that handles missing values, corrects invalid entries, and prepares clinical datasets for analysis.	https://dx.doi.org/10.6084/m9.figshare.32146693
MLSight	A tool to build and evaluate machine learning models, allowing users to train algorithms and assess performance metrics.	https://dx.doi.org/10.6084/m9.figshare.32146696
ValidateSight	A tool for the validation of machine learning models that can be used for external validation	https://dx.doi.org/10.6084/m9.figshare.32727849
PredictSight	A deployment tool that uses trained models to generate predictions for new patient data in real-time clinical settings.	https://dx.doi.org/10.6084/m9.figshare.32146699
R	A programming language used for statistical computing, data analysis, and machine learning.	https://cran.r-project.org
RStudio	An integrated development environment (IDE) that makes it easier to write, run, and manage R code.	https://posit.co/download/rstudio-desktop
Example data set	This is a fictitious dataset (300 patients) containing demographics, history, anthropometric, laboratory values, and presence and absence of nonalcoholic fatty liver disease (17 parameters)	https://dx.doi.org/10.6084/m9.figshare.32146681

Note: CleanSight, MLSight, PredictSight, and the example dataset are prepared by the authors.

Case Study: Building a Prediction Model for Nonalcoholic Fatty Liver Disease

To demonstrate the practical application of a no-code ML workflow, we developed a case study using a fictitious dataset designed to simulate real-world clinical scenarios. ³⁵ The dataset includes common clinical and biochemical parameters such as age, body mass index (BMI), blood glucose, and lipid profile, which are relevant for predicting the risk of nonalcoholic fatty liver disease. To reflect routine clinical data challenges, the dataset intentionally includes missing values and minor inconsistencies.

Preparing the Computer Environment

The applications were developed using R and the Shiny framework and are compatible with 64-bit Windows, macOS, and Linux operating systems. A computer with a dual-core processor, 4 GB of random access memory (RAM), and approximately 2 GB of free disk space is sufficient for basic operation; however, a system with an Intel Core i5 or AMD Ryzen 5 processor (or equivalent), 8 GB of RAM, and a solid-state drive is recommended to ensure smooth performance during data processing and ML model development. For large datasets (e.g., more than 50,000 observations) or computationally intensive algorithms such as random forest, gradient boosting machine, and extreme gradient boosting (XGBoost), 16 GB or more of RAM is advisable. The Shiny applications are optimized for CPU-based computation, making them suitable for use on standard desktop and laptop computers commonly available in academic and health care institutions.

Users need to install R and RStudio, and the computer should be connected to the internet before starting the following steps. Shiny applications run on R and its packages. ³⁶ When the code is entered for the first time, the required packages will be downloaded for functionality. Hence, an internet connection is required, and it will take some time. After the installation, next time onwards, when the packages are saved, they will not require an internet connection. After a one-time installation and initial script launch, routine use of the application does not require users to write any code. To start the application, users simply need to copy the script from the provided Word document, paste it into the R console, and run it. Hence, the term “no-code” is used in this article.

To use CleanSight, MLSight, ValidateSight, and PredictSight, users first need to download the application files and save them on their computer for future use. Each application is provided as an R script that can be opened in Microsoft Word. After launching RStudio, the user opens the downloaded code, pastes it into the RStudio console ( Fig. 1 ), and presses Enter or Return. During the first execution, RStudio automatically installs any required R packages that are not already available on the system. Once the installation is complete, the application launches in a web browser or in RStudio as an interactive Shiny interface, where users can upload datasets, configure the required options, and perform the desired analyses without writing any additional R code.

Fig. 1.

Console of RStudio where users paste the code to run Shiny applications. The screenshot shows the RStudio console with intentional placeholder text “PASTE CODE HERE” to indicate that users need to paste the code of CleanSight, MLSight, and PredictSight immediately after the “ > ” symbol.

Step 1: Data Cleaning with CleanSight

The first step involves preparing the dataset using CleanSight. The user uploads the raw dataset in the Microsoft Excel (.xlsx) format containing the clinical variables. After uploading, the dataset is divided into training and testing datasets according to a user-defined proportion (e.g., 80% training and 20% testing), ensuring that all subsequent preprocessing parameters are learned only from the training data. The user interface of CleanSight is shown in Fig. 2 . The application then performs sequential data preprocessing by imputing missing numerical values using a user-selected method (e.g., median), with optional conversion of biologically implausible negative clinical values to zero and encoding of categorical text variables into numerical values when required. Importantly, the imputation parameters estimated from the training dataset are subsequently applied to the testing dataset to prevent information leakage and maintain methodological rigor. Both the raw split datasets and the cleaned/imputed training and testing datasets can be downloaded separately for subsequent ML model development in MLSight.

Fig. 2.

User interface of the CleanSight application showing the workflow for data preprocessing, including dataset upload, training–testing data splitting, missing value imputation, optional handling of negative values and categorical variables, and downloading of the processed training and testing datasets for subsequent machine learning model development. The screenshot shows the configuration panel of CleanSight, where a dataset (fatty_liver_dataset.xlsx) has been uploaded successfully. The cleaned dataset can then be downloaded using the “Download Imputed Train” and “Download Imputed Test” options.

Users can download the example dataset ( Table 4 ) and clean the data according to necessity.

Step 2: Model Building and Evaluation with MLSight

After the training and testing datasets have been prepared using CleanSight, the application is closed and MLSight is opened. The cleaned training and testing datasets are uploaded separately into MLSight ( Fig. 3 ). The user then selects the outcome (target) variable that the model should predict and chooses a ML algorithm from the available options.

Fig. 3.

User interface of the MLSight application showing model configuration options, including training and testing dataset upload, target variable selection, machine learning algorithm selection, class imbalance handling, hyperparameter tuning, probability calibration, and cross-validation settings before model training. Cleaned dataset (training and test) has been successfully uploaded. The user selects the target variable (fatty liver), defines the task type as classification, and chooses the algorithm (Random Forest). There is an option to select class imbalance handling, enabling hyperparameter tuning and calibration. The execution pipeline includes three sequential steps: training the model, testing its performance, and downloading the trained model file (not shown in the figure).

In ML, the first and most important step is to clearly define the task type, because it determines the entire modeling strategy. Broadly, clinical prediction problems fall into three categories. Classification is used when the outcome is categorical, such as predicting whether a patient has fatty liver disease (yes or no). Regression is used when the outcome is a continuous number, such as blood pressure or liver fat percentage. ³⁷

Once the task is defined, the next step is selecting an appropriate algorithm. This can be thought of as choosing a statistical “tool” for the problem. Simple and highly interpretable methods such as logistic regression are often used as a starting point, especially in clinical settings where transparency is important. More flexible methods like random forests are useful when relationships between variables are complex and nonlinear. Common models are shown in Table 2 . ³⁸

A major challenge in clinical datasets is class imbalance, which occurs when one outcome category is much more frequent than the other. ³⁹ For example, if 95% of patients do not have a disease and only 5% do, a model could incorrectly appear accurate by simply predicting “no disease” for everyone. Class imbalance can be mild, moderate, or severe, depending on how unequal the distribution is. When dealing with class imbalance in clinical datasets, several resampling techniques are commonly used to improve model learning. Upsampling involves increasing the number of cases in the minority class by randomly duplicating existing observations so that both classes become more balanced. While simple, it can sometimes lead to overfitting because the model sees repeated identical cases. Downsampling works in the opposite way by reducing the number of cases in the majority class, creating balance by removing some observations. This can help speed up model training, but may discard potentially useful information. A more advanced method is SMOTE (Synthetic Minority Over-sampling Technique), ⁴⁰ which improves upon simple upsampling by creating new, synthetic minority class examples rather than duplicating existing ones. It does this by generating new data points that lie between existing minority class samples, making the dataset more diverse and reducing the risk of overfitting. In practice, SMOTE is often preferred when the dataset is small to moderate, and the minority class is underrepresented.

Another important concept is hyperparameter tuning, which refers to optimizing the settings that control how a ML model learns. ⁴¹ Unlike model parameters, which are learned from data, hyperparameters are set before training begins. Examples include the number of trees in a random forest, the maximum depth of trees, or the learning rate in boosting algorithms. Tuning involves systematically trying different combinations of these settings to identify the configuration that produces the best performance.

In many clinical applications, it is not enough for a model to simply rank patients correctly; it must also provide accurate probabilities. This is where Platt calibration becomes important. ⁴² Some ML models produce probability estimates that are poorly calibrated, meaning that predicted risks do not match actual observed outcomes. For example, a model might predict a 90% risk of disease, but only 70% of such patients truly have the disease. Platt calibration adjusts these outputs so that predicted probabilities better reflect real-world frequencies, making them more trustworthy for clinical decision-making.

In addition to this, MLSight provides the option of cross-validation. In this approach, the dataset is divided into multiple parts, and the model is trained and tested repeatedly on different combinations of these parts. Each repetition provides an estimate of performance, and the final result is usually the average across all repetitions. A key concept here is the fold, which refers to each subset of the data used in cross-validation. For example, in fivefold cross-validation, the data are split into five equal parts; the model is trained on four parts and tested on the remaining one, and this process is repeated five times so that each part serves once as the test set. ⁴³ Fivefold cross-validation is commonly used because it provides a good balance between reliability and computational efficiency, while 10-fold cross-validation may be preferred in smaller datasets for more stable estimates.

After selecting the options in MLSight, clicking on the “Execute Training Protocol” will train the model with the selected data, and after that, clicking on “Evaluate Unseen Cohort” will test the performance of the model in terms of the metrics shown in Table 3 . ²⁸

For the example data, the trained model was evaluated using standard classification performance metrics, including accuracy, area under the receiver operating characteristic (AUC-ROC) curve, F1 score, precision, sensitivity, and specificity ( Fig. 4 ). Its ability to discriminate between positive and negative outcomes was further assessed using the ROC curve ( Fig. 5 ). Model calibration was then examined by comparing predicted probabilities with observed outcomes to evaluate the reliability of risk estimates ( Fig. 6 ). Finally, the relative importance of individual predictor variables was analyzed to identify the factors contributing most to the model's predictions ( Fig. 7 ). The displayed results are illustrative only, should not be interpreted as performance benchmarks, and are dependent on the dataset, model selection, and training parameters.

Fig. 4.

Performance metrics of the trained machine learning model. Bar chart summarizing the model performance on the test dataset, including accuracy, AUC-ROC, F1 score, precision, sensitivity (recall), and specificity. Synthetic demonstration only, not for clinical interpretation. AUC-ROC, area under the receiver operating characteristic curve.

Fig. 5.

Receiver operating characteristic (ROC) curve of the trained model. ROC curve showing the discrimination performance of the prediction model, with the reference line representing random classification. Synthetic demonstration only, not for clinical interpretation.

Fig. 6.

Calibration plot of the prediction model. Calibration plot comparing predicted probabilities with observed event frequencies. The diagonal line represents perfect calibration. Synthetic demonstration only, not for clinical interpretation.

Fig. 7.

Feature importance of the trained machine learning model. Variable importance plot showing the relative contribution of predictor variables to the model, ranked from highest to lowest importance. Synthetic demonstration only, not for clinical interpretation.

The trained model can then be saved as a file (e.g., .rds), representing the “learned brain” of the system.

Step 3: External Validation with ValidateSight

In the domain of clinical AI, a model that performs flawlessly during internal testing can fail dramatically when deployed in a real-world hospital. ⁴⁴ External validation by ValidateSight facilitates evaluation on a user-supplied external dataset. The validity of the assessment depends on the independence, quality, sample size, and representativeness of that dataset.

It begins with model ingestion, where the finalized model archetype bundle (.rds) exported from MLSight is loaded into ValidateSight ( Fig. 8 ). Next, a completely independent cohort dataset from a separate institution or time period is uploaded. Before evaluation, the application executes an automated integrity audit, generating a predictor match matrix that verifies whether the new dataset contains all necessary variables, matches the original data types, and aligns perfectly with expected factor levels to prevent catastrophic runtime failures.

Fig. 8.

User interface of ValidateSight showing the option to upload a model, external data, and test the model. The audit configurations panel allows the user to upload a previously trained machine learning model architecture (.rds) and an independent external validation dataset (.xlsx or .csv), followed by selection of the target outcome variable. After configuration, the Execute External Validation Protocol button initiates the automated external validation process. Upon successful completion, ValidateSight confirms that the external audit has been completed and prompts the user to inspect the generated verification sheets containing the validation results. Synthetic demonstration only, not for clinical interpretation.

Once structural integrity is confirmed, ValidateSight evaluates the model across two fundamental pillars of clinical performance: discrimination and calibration. The AUC-ROC, along with its 95% confidence intervals, ⁴⁵ ensures the model's diagnostic boundaries remain statistically stable. Alongside the AUC, traditional classification metrics like sensitivity, specificity, and F1 are also provided.

Simultaneously, the system assesses clinical safety through probability calibration, ensuring that the model's risk percentages mirror true clinical frequencies. This is measured via the Brier score, which quantifies the mean squared error of the probabilities on a scale where zero represents absolute perfection.

Step 4: Clinical Prediction with PredictSight

In the final step, the saved model is uploaded into PredictSight ( Fig. 9 ). This tool allows users to input new patient data, including clinical and laboratory parameters, and generate a prediction of risk in real time. The input parameters will automatically appear according to the trained model data. The output can be visualized on the screen and can be downloaded. This stage represents the practical deployment of the model, transforming previously collected data into a usable decision-support tool.

Fig. 9.

User interface of PredictSight showing system setup, patient identifier, and clinical measurements as predictor variables. The screenshot displays the PredictSight interface, where the trained model can be uploaded under System Setup. The patient identifier is a generic input field, and all the fields of predictor variables will appear under Clinical Measurement according to the uploaded trained model. Users can generate a prediction and can download a PDF report for future reference. Synthetic demonstration only, not for clinical interpretation.

For the above example, data fields such as age, sex, height, weight, BMI, etc. will appear under Clinical Measurements. After filling in the details and clicking on “Generate Prediction,” the result will display as either “Yes – High risk” or “No – Low risk.”

An example result in PredictSight is shown in Fig. 10 .

Fig. 10.

A result in PredictSight for patient screening. PredictSight interface showing model upload, patient data entry, and real-time prediction output with the generated diagnostic result. Synthetic demonstration only, not for clinical interpretation.

All the data, results, and ML models are fictitious and presented here only for demonstration. Please do not interpret any result for clinical use. For example, with a sensitivity of approximately 0.85 (85%) and a specificity of approximately 0.32 (32%), the model is not suitable as a standalone screening or diagnostic tool, primarily due to its very low specificity. Screening tests generally require high sensitivity and reasonable specificity to avoid overwhelming the health care system with false positives. Diagnostic tests require both high sensitivity and high specificity, as treatment decisions are based on their results.

Discussion

The present review addresses a critical gap between the growing availability of clinical data and the limited ability of health care professionals to utilize these data for predictive modeling. This article contributes meaningfully by reframing ML not as a complex computational discipline reserved for data scientists, but as an accessible, structured workflow that clinicians can engage with using intuitive tools. With this guidance, busy physicians would learn the basics of ML, use their data to build a model, and actually deploy it.

We chose R and its Shiny framework because R is a popular language for statistics, ⁴⁶ and Shiny lets us wrap complex code into simple, point-and-click interfaces. This framework removes all that technical friction. All the applications are open source, and data remain on personal computers. Local processing may reduce the need to transmit data to external servers, but it does not by itself ensure data security. While the current framework appropriately highlights local data processing on personal computers, it is important to explicitly mention cybersecurity safeguards. Even in local environments, data protection measures such as secure storage practices and institutional governance oversight are essential to prevent unauthorized access or data breaches. ⁴⁷

A brief cautionary note is warranted regarding the limitations and responsible use of ML models in clinical contexts. The model performance is highly dependent on the quality, representativeness, and completeness of the underlying dataset; therefore, biased, incomplete, or low-quality data may lead to misleading or nongeneralizable predictions. ⁴⁸ Inappropriate deployment of such models in clinical practice without rigorous external validation may pose risks to patient safety.

Additionally, issues related to model interpretability and explainability remain critical, particularly in health care settings where decision transparency is essential for clinician trust and ethical accountability. Regulatory and ethical considerations, including compliance with institutional policies and relevant data protection frameworks, must be addressed prior to real-world implementation. ⁴⁹ In addition, a model once made cannot be used for a long time. Continuous model monitoring after deployment, including drift detection (identifying shifts in patient data patterns or model performance over time that can degrade predictive accuracy), periodic retraining, and performance re-evaluation, ensures sustained accuracy over time. ⁵⁰ Integration into hospital information systems should be planned carefully to support workflow compatibility and clinical usability.

While building ML and publishing it, authors should follow any standard guidelines. Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis-Artificial Intelligence (TRIPOD-AI) extends the original TRIPOD framework by providing specific recommendations for reporting ML-based prediction model studies, including clear documentation of data preprocessing, model development, validation strategy, performance metrics, and handling of missing data. ⁵¹ Adherence to these guidelines enhances reproducibility, facilitates critical appraisal, and ensures that key methodological details necessary for external validation and clinical translation are adequately reported.

Although several no-code ML platforms, such as WEKA, Orange, and Google AutoML, are available, they are primarily designed as general-purpose ML tools rather than clinician-oriented applications. Google AutoML automates model development but is a cloud-based commercial platform that may require internet connectivity, cloud infrastructure, and technical configuration, making it less suitable for institutions with strict data privacy requirements. In contrast, our platform was specifically developed for clinicians and biomedical researchers with minimal programming experience. All source code is fully open source, allowing users to inspect, modify, validate, and extend the software, thereby enhancing transparency, reproducibility, and adaptability for research and clinical applications.

The article primarily focuses on structured datasets and basic ML models, which may not fully capture the complexity of modern health care data, such as imaging, genomics, or unstructured clinical text. Additionally, while the no-code approach improves accessibility, it may limit users' understanding of underlying algorithms and reduce transparency in model behavior. In addition, we did not include any image analysis model that requires deep learning and a higher configuration in a computer, and more time to build the model. Hence, this article has more pedagogical value, which would help clinicians understand the basics of ML.

Conclusion

AI has the potential to transform health care, but its technical complexity has historically limited direct engagement by many health care professionals. This article outlines a practical approach for physicians and other clinicians with limited ML expertise to understand and explore structured-data ML workflows using no-code platforms. However, caution is warranted, as models intended for clinical use require rigorous methodological review, independent validation, governance approval, and prospective evaluation before they can be safely implemented in clinical practice.

How to Use these Applications in Research and Publication

The applications (CleanSight, MLSight, ValidateSight, and PredictSight) are open source under the Massachusetts Institute of Technology License, a permissive open-source license that allows users to use, modify, and redistribute the software without requiring permission. Researchers who use these applications to build or compare models and publish their work are encouraged to cite this article for attribution.