As of 2026, I no longer maintain a public library of the eHealth tools in my portfolio. Below are descriptions of the tools I have designed, deployed and evaluated over the past 15 years, from early adaptive tests to the large language model (LLM) and agentic systems I work on today.
Many of the earlier tools were built with Concerto, the University of Cambridge Psychometrics Centre’s open-source platform for electronic assessments and web-based statistical applications. Where demos still exist, they run on servers that receive only occasional maintenance, so some links may not work.
AI at scale: Oracle Health
At Oracle Health, my teams have taken more than twenty AI systems from idea to general release, spanning predictive, causal, generative and agentic models. They help health systems and life sciences partners find eligible patients for clinical trials, monitor drug safety, summarize long patient records and check care against clinical guidelines. Building them taught me that the hard part is not making an AI tool but knowing whether it works. We therefore developed new evaluation frameworks for generative and agentic AI, combining “agent-as-judge” methods, structured benchmarks and continuous monitoring after go-live. These products are proprietary, so there are no public demos, but I am always happy to talk about how we evaluate them.
ResPECT: LLM-based assessment for the Lancet Commission on Cancer and Health Systems
Most people affected by cancer are never asked about the full impact it has on their lives. As data science and AI lead for the Lancet Commission on Cancer and Health Systems, I developed ResPECT, an assessment that uses LLMs to turn people’s own words about their cancer experience, in more than 50 languages, into structured insights for policymakers. We built a companion dashboard with the World Health Organization’s International Agency for Research on Cancer (IARC), and the European Commission has awarded $20 million to scale the work across Europe.
Fair, safe and interpretable prediction tools (with Project Ronin)
A model that performs well in the lab can still fail patients at the bedside. Working with Project Ronin, we deployed machine learning tools that predict risks such as unplanned acute care, built in checks for fairness, safety and interpretability, and published how we did it.
PatientMosaic at MD Anderson
PatientMosaic aims to collect diverse, longitudinal multiomic data, including patient-reported outcomes (PROs), from 10,000 patients at MD Anderson. To help meet this goal, we developed a next-generation patient assessment tool that combines adaptive testing algorithms, fuzzy search, and individualized feedback and guidance from ASCO. Compared with systems without these features, patient retention through the study improved by 50%, and we captured 10% more information about patients’ symptoms than static assessments did.
Immunotherapy Toxicity (IOTOX) Guidelines at MD Anderson
Immunotherapies have transformed the treatment of many cancers, but they can cause severe side effects. We drew on the Division of Internal Medicine’s deep experience in treating these rare toxicities to build a dynamic, electronic IOTOX treatment guideline system.
ATLanTiC Quality of Life Assessment
ATLanTiC is a computerized adaptive test developed with the WHO International Hub for Quality of Life Research at the University of Manchester. It chooses each question based on the user’s previous answers, so assessments are shorter and more reliable than paper questionnaires. You can read the accompanying journal article here.
Quality of life prediction tool
It is true that “all CATs are black in the dark”: once computerized adaptive tests are deployed in the real world, it is hard to judge how well they are working because we lose the ground truth. This system lets users rate the model’s predictions about their quality of life, which helps validate the models. I received the 2016 New Investigator Award at the 23rd Annual Conference of the International Society for Quality of Life Research (ISOQOL) for presenting this method.
GMC Colleague Questionnaire open-text analysis
Many health services collect open-text feedback, but it often goes unused. When qualitative data arrive in enormous quantities, human raters are a slow and expensive way to make sense of them. We trained an ensemble of machine learning algorithms that classified open-text reports of doctors’ performance with startling accuracy. This was an early step toward the language models I work with today.
Computerized adaptive depression assessment using the CES-D
An adaptive version of a popular “legacy” depression screening questionnaire, developed with Aiden Loe, then a doctoral candidate in psychology at the University of Cambridge.
Machine Interpretation of Natural Text – Sentiment Analysis in the Concerto Environment (MINT-SAUCE)
An example of machine learning sentiment analysis built in Concerto, which I used in my machine learning teaching.
Other tools developed elsewhere
Predictive World (Gibbons, Hoferkova, Barclay and Popov, with Sid Lee, Stink Digital and Ubisoft Entertainment)
The Psychometrics Centre worked with industry partners on a predictive online system to support the launch of Watch Dogs 2. Predictive World draws together 2.5 billion pieces of publicly available data to make predictions about individuals, now and in the future. More than a marketing tool, it draws on epidemiological and psychological research so users can explore how lifestyle choices such as smoking and drinking affect outcomes in health, work and leisure. It attracted more than a million users in its first two weeks, was well received internationally, and won two prestigious design awards: the D&AD Graphite Pencil and the ADC Gold Cube.
Apply Magic Sauce (Kosinski, Stillwell and Rust)
Apply Magic Sauce (AMS, to its friends) is a personalization engine that predicts psychological traits from digital footprints of human behavior. It uses machine learning to predict intimate psychological traits from Facebook “Likes” (papers here and here) and open text.