Artificial intelligence in real-world evidence
Artificial intelligence and machine learning can clean, join and analyse real-world data, and they predict outcomes and find patterns in it. Each common algorithm gives its own kinds of output, biased training data and opaque models are concerns in using them, and the medical writer explains the model, its outputs and its limits to the people who act on them.
What the tools do with real-world data
AI and machine learning can analyse large and complex datasets and identify patterns and insights that were previously hidden or too difficult to detect with traditional analytical methods. Integrated into real-world evidence, they enable predictive modelling, risk stratification and clinical decision support. They improve data quality by automating error detection, managing missing values and standardising formats. They analyse unstructured text, images and audio through natural language processing and computer vision. To integrate data, they resolve entity mismatches and map datasets to common frameworks, and federated learning trains models without sharing raw data. They forecast trends and recommend actions such as personalised treatments, and they process streaming data in real time, with edge computing enabling instant analysis on devices. They detect bias in datasets and design fairness-aware algorithms, and synthetic data help reduce bias and protect privacy. Automated machine learning and distributed computing let them scale, federated learning and differential privacy preserve privacy, and explainable AI makes models transparent.
Applied to large real-world datasets, AI and machine learning can forecast patient outcomes and the progression of disease. Predictive models can estimate the likelihood of complications, hospital readmission or treatment response, so that interventions can be tailored in advance, and this predictive capability helps identify at-risk patients who may benefit from early intervention. Models of disease progression let clinicians identify high-risk patients who may need more aggressive interventions. In diabetes, for example, algorithms can detect patterns that point towards retinopathy, nephropathy or cardiovascular disease from trends in haemoglobin A1c, blood pressure and renal function markers. In heart disease, models can weigh cholesterol, previous cardiac events, medication adherence and electrocardiogram findings to estimate the likelihood of heart failure exacerbation or myocardial infarction.
The same tools can track how treatment approaches evolve, identify the most effective therapies in practice and reveal how care differs across patient groups. Machine learning can find the most frequently prescribed treatments for a condition in health records, so that clinicians and policymakers can evaluate the effectiveness of current practice, and such findings can support cost-effectiveness analysis. Real-world data on admissions and discharges can show how effective post-surgical care plans are, and readmission rates and length of stay let hospitals judge whether patients receive adequate care after discharge and find the risk factors that lead to complications. Identifying common causes of readmission, such as infections or medication non-adherence, allows targeted interventions that improve the coordination of care and prevent unnecessary hospitalisation. Predictive analytics can further improve the use of resources by optimising staffing, bed management and follow-up care.
Further applications are:
- genomic analysis for targeted cancer therapy
- forecasting which patients are at high risk of hospital readmission from comorbidities, past stays, laboratory results and social determinants in health records
- post-marketing pharmacovigilance that detects adverse events and safety signals in near real time from health records and claims
- stratifying patients into risk categories and forecasting disease trajectory
Generative AI in trial design
Generative AI helps enhance the design of a trial, such as its inclusion and exclusion criteria, endpoints and comparator arms, from real-world data such as health records, claims and registries, leading to faster recruitment, lower cost, dynamic protocol adjustment and higher external validity. It ingests and harmonises unstructured data such as clinician notes and laboratory results, and it simulates patient cohorts that reflect the comorbidities, treatment patterns and geography of real populations. It generates synthetic control arms from matched historical patient data, which reduce or replace the need for placebo groups, assess the comparative effectiveness of drugs and aid endpoint selection. For feasibility and site selection, it predicts patient availability by site and region and simulates synthetic populations for pretrial feasibility assessment. It has been used substantially to generate and refine trial protocols by predicting dropout rates, identifying high-performing sites and assessing protocol feasibility from real-world data.
The algorithms and what they output
Supervised, unsupervised and reinforcement learning are the core machine learning approaches applied in healthcare and real-world evidence generation. Supervised models such as logistic regression, decision trees, random forests and XGBoost are commonly applied in diagnosing conditions such as diabetes or hypertension. Unsupervised techniques help group patients or detect anomalies without predefined outcomes. Reinforcement learning is promising for dynamic treatment decision support, despite challenges in defining rewards and causal pathways.
- Logistic regression models the probability of a binary event from predictor variables. It is used to predict clinical outcomes and adverse events and to stratify patients by risk, and the probability score it calculates classifies patients as high or low risk, which helps prioritise those needing immediate care. Performance metrics such as accuracy, precision, recall, the F1 score, which is the harmonic mean of precision and recall, and the area under the ROC curve assess how well it classifies patients. Its coefficients show how much each factor contributes to a patient's risk level, a positive coefficient indicating a higher likelihood of being high-risk, and the confusion matrix and ROC curve show how well the model tells high-risk from low-risk patients, or slow from fast progressors.
- A decision tree, effective for classification and regression problems, segments the population hierarchically from a root node through branches and decision nodes to leaf nodes that represent the outcome. Each decision helps categorise a patient as high or low risk. For example, a decision tree was used to predict the risk of cardiovascular events in a large patient dataset from factors such as blood pressure and cholesterol. Its key strength is interpretability: it shows the exact conditions behind each classification, which helps identify patients likely to move to a higher risk category. It also reports feature importance, which shows the variables with the most impact on its predictions. A decision tree may do better than logistic regression because it captures non-linear relationships and handles mixed data types. Logistic regression assumes a linear relationship between the predictors and the outcome, while a decision tree can model complex interactions between variables without assumptions about the data distribution, which lets it predict more accurately where the relationships between variables are not straightforward.
- A random forest builds many trees on bootstrap samples of the data and of the features at each split and averages their predictions. This helps reduce the overfitting a single tree is prone to and improve generalisation to new data. A random forest can handle complex interactions and make better predictions, especially where logistic regression or a single tree does not give satisfactory results, and it reports feature importance. In one example a random forest predicted phenotype transformations from genetic and environmental interactions.
- XGBoost builds trees sequentially, each correcting the errors of the one before on misclassified data points, and uses regularisation against overfitting. It handles missing data, has highly parallelisable code for large and complex datasets, and uses a sparsity-aware algorithm and a weighted quantile sketch for approximate tree learning. It can analyse genetic factors, medical history and lifestyle to predict transition to an advanced disease stage, and it reports feature importance. In one study predicting the progression of chronic kidney disease, it outperformed other models at identifying patients at high risk of rapid progression.
- Unsupervised algorithms detect clusters, reduce dimensionality and uncover latent structure, and they are increasingly used in health economics and outcomes research, pharmacovigilance and post-market surveillance. Clustering such as k-means and hierarchical clustering groups similar patients on clinical and demographic variables, and dimensionality reduction such as principal component analysis and t-SNE simplifies high-dimensional data. Their outputs are cluster assignments, group labels, visualisations such as heatmaps, dendrograms and scatter plots, and metrics such as the silhouette score, variance explained and cluster centroids. In one example, clustering found three subtypes of chronic kidney disease, and one subgroup, 30% of the population, showed frequent treatment switching and more hospitalisation, which indicated a high-risk phenotype with possible unmet needs.
Bias and opacity
Ensuring that training data are unbiased is a major challenge in applying AI and machine learning to real-world evidence. Real-world data often under-represent some populations, which can skew outcomes and cause disparities in healthcare, and models trained on such data may reinforce these biases and predict less accurately for the under-represented groups. Approaches that could resolve this are diverse, representative training data, regular audits of algorithms and fairness-aware techniques.
Many AI systems, especially those based on deep learning, do not show how a decision was reached, although clinicians, patients and regulators need to see the reasoning behind the outputs. Approaches that could foster transparency are explainable AI such as feature importance analysis and decision trees, interfaces that convey outputs clearly, documentation and training on assumptions, limitations and appropriate use, and collaboration between data scientists, clinicians and regulators so that models are technically robust and meet clinical needs.
The writer's role
Communicating complex, algorithm-driven insights to the various healthcare stakeholders is a critical challenge. For medical writers involved in regulatory submissions, clinical research documentation and healthcare communications, understanding the application of AI and machine learning in real-world evidence generation is essential. Writers need not be data scientists, but they must understand the fundamentals of analytic methods. Regulators, clinicians, pharmaceutical companies and policymakers need clear, precise and scientifically accurate interpretations of AI-generated evidence, and writers act as the intermediaries who make those interpretations methodologically sound, comprehensible, compliant and aligned with decision-making frameworks. The key responsibilities in translating the results of such models are:
- bridging the language gap, by translating between data scientists and healthcare audiences and interpreting model assumptions, methodology, data inputs, limitations and outputs such as risk scores, clusters and probabilities
- identifying and emphasising the results that reveal clinically significant subgroups, suggest differences in treatment response, or indicate the burden of disease or health outcomes
- crafting clinically meaningful narratives, as in "The model identified a patient segment at threefold higher risk of hospitalization within 12 months, characterised by polypharmacy, diabetes, and advanced age. This suggests a target for early intervention."
- supporting visual interpretation, by developing or adapting visualisations such as Kaplan–Meier curves or cluster heatmaps, annotating clinically relevant findings and making figures publication- or submission-ready
- ensuring scientific and regulatory rigour, which in real-world evidence deliverables, especially for HTA or regulatory use, means writers must clearly describe the model type, inputs and limitations, not over-interpret exploratory analyses, and frame findings within accepted standards
- presenting the value of the techniques in peer-reviewed manuscripts and conference presentations
Writers translate outputs such as predictive models, risk scores and pattern recognition into actionable insight for clinicians, focused on improving patient outcomes or streamlining clinical workflows. They explain data sources, methods and limitations in a way that is both scientifically rigorous and easy to understand, and by providing context and highlighting the most relevant takeaways they help clinicians and stakeholders trust and use the findings in decision-making. They avoid overhyped claims, stress the need for responsible interpretation and address the limitations of the data and models. As AI-derived real-world evidence becomes increasingly central to healthcare innovation, the writer's role is evolving from content developer to strategic partner. One study that used AI on health records to predict cardiovascular outcomes required writers to translate the machine learning models into actionable insight for clinicians, and in another case writers ensured that evidence extracted by natural language processing from unstructured clinical notes was accurately represented in regulatory submissions. By collaborating with data scientists, clinicians and regulators, writers can help unlock the potential of AI in healthcare.