Medicine has long been a practice of reading the past to anticipate what comes next, and a new tool called Foresight extends that old habit into the territory of machines. Built by a team including investigators from King's College London and trained on the health records of more than 811,000 NHS patients, it learns from the accumulated traces of care to estimate the course a person's illness may take. Its makers offer it as an aid to the clinician's judgment rather than a voice for the patient, and whether it earns a place at the bedside will depend on tests that have yet to be run in real ho
King's College London Team Unveils Foresight, a Deep Learning Tool That Predicts Patient Health Trajectories
Not to predict your future, but to make sure a diagnosis isn't missed.
Mark
Start with the basics. What does Foresight actually do?
Mimi
It reads patient records and predicts what is likely to come next, whether that's a disorder, a symptom, a relapse, or a medication. It was trained on data from more than 811,000 NHS patients.
Mark
And it reads free text too, not just tidy fields?
Mimi
Yes. Age, ethnicity, and sex are the structured parts, but it also processes the free-text notes, which is where a lot of clinical detail lives.
Luke
Let's be careful with the numbers. The headline accuracy is 68% and 76% at two NHS Trusts, and 88% on the US MIMIC-III dataset. That's a wide spread.
Mimi
It is, and the source doesn't explain why the US figure is higher. The task was forecasting the next disorder among 10 candidates.
Mark
What about the 93% figure?
Luke
That one is narrower than it sounds. It applies to mock timelines clinicians built, and only where they unanimously agreed on a predicted event. It isn't 93% accuracy on real patients.
Mimi
Fair. It measures clinical relevance in constructed scenarios.
Mark
So what is it meant for?
Mimi
Kraljevic says it's a clinician aid, for catching missed diagnoses or continual monitoring, not for patients to self-diagnose.
Luke
And his claim that it scales easily and improves with more data is the developers' own statement. Nothing in the reporting shows it tested in live care, so whether it improves outcomes is still unknown.
Il Polso
- A deep learning model trained on more than 811,000 NHS patient records is being positioned to anticipate a patient's next medical event.
- Its strongest results come from retrospective comparisons and clinician-built mock scenarios, leaving its performance in live care untested.
- Researchers are steering the tool firmly toward clinicians, explicitly warning against patient self-diagnosis and self-prediction.
- Its developers see potential in continuous monitoring, missed-diagnosis prevention, trial design, and education, but real-world validation and clinical trials are still required.
Medicine has long been a practice of reading the past to anticipate what comes next, and a new tool called Foresight extends that old habit into the territory of machines. Built by a team including investigators from King's College London and trained on the health records of more than 811,000 NHS patients, it learns from the accumulated traces of care to estimate the course a person's illness may take. Its makers offer it as an aid to the clinician's judgment rather than a voice for the patient, and whether it earns a place at the bedside will depend on tests that have yet to be run in real hospitals.
A new artificial intelligence tool called Foresight, developed by a research team that includes investigators from King's College London, has been trained on the electronic health records of more than 811,000 NHS patients. Its purpose is to forecast what is likely to happen next in a patient's medical history, and its creators hope it could eventually support clinical decisions, healthcare monitoring, and the running of clinical trials.
The model uses deep learning to find patterns across two kinds of information: structured details such as age, ethnicity, and sex, and the free-text notes clinicians write during care. Using these records, the team trained three distinct Foresight models and tested them against outcomes already recorded. Foresight correctly identified the next disorder in a patient's timeline 68% and 76% of the time at two UK NHS Trusts, and 88% of the time on the US MIMIC-III dataset.
Clinicians also reviewed the tool using mock patient timelines across varied medical scenarios. Where they agreed on a predicted event, they judged Foresight's predictions 93% relevant from a clinical standpoint.
Zeljko Kraljevic, a research fellow in health informatics at King's College London, was clear about the tool's limits. It is not meant for patients to diagnose themselves or forecast their own futures. Instead, clinicians might use it to avoid missing a diagnosis or to monitor patients continuously with real-time risk estimates. He also highlighted scale as a strength, noting the system can extend to more patients, hospitals, or disorders with little modification and can improve as it receives more data.
For now, these results come from comparisons with past records and from simulated cases, not from live clinical use. Whether Foresight truly helps doctors in practice remains an open question, to be answered by real-world validation and clinical testing.
Foresight, a new artificial intelligence tool built by a research team that includes investigators from King's College London, has been trained on the electronic health records of more than 811,000 patients in the NHS. Its aim is to forecast what is likely to happen next in a patient's medical history. The researchers say it could eventually support clinical decision-making, healthcare monitoring, and the efficiency of clinical trials.
The tool uses a deep learning approach to find complex patterns in the records, which hold both structured and unstructured information. The structured data covers details such as age, ethnicity, and sex. The unstructured data is free text, the kind of notes clinicians write in the course of care. The team extracted and processed both kinds of data, and used the records to train three distinct Foresight models.
To test whether the predictions held up, the researchers compared them with the health outcomes actually recorded in a smaller subset of data. In one test, Foresight was asked to forecast the next 10 possible disorders to appear in a patient's timeline. It correctly identified the next disorder 68% and 76% of the time at two UK NHS Trusts, and 88% of the time on MIMIC-III, a US dataset.
A second test asked the model to predict the next new biomedical concept, which could be a disorder, a symptom, a relapse, or a medication. Here Foresight reached precision of 80%, 81%, and 91%, respectively.
Clinicians also assessed the tool directly. They built mock patient timelines covering a variety of medical scenarios. Where the clinicians were unanimous about a predicted medical event, Foresight's predictions were judged 93% relevant from a clinical standpoint. The team points to this as evidence of potential uses in risk forecasting, emulating clinical research, studying how disorders progress, simulating interventions, and education.
Zeljko Kraljevic, Research Fellow in Health Informatics at King's College London, said the study shows Foresight can predict patients' health trajectories with high precision, and could be a valuable aid to decision-making and clinical research. He was explicit about what the tool is not for. It is not intended to let patients diagnose themselves or predict their own futures. Instead, he said, clinicians might use it to make sure a diagnosis is not missed, or for continual monitoring that gives real-time risk predictions.
Kraljevic also described scale as a main advantage. According to him, Foresight can extend to more patients, more hospitals, or more disorders with minimal or no modification, and it improves as it receives more data.
The figures so far come from comparisons with recorded data and from clinician-built mock scenarios, not from use in live care. Whether the tool helps clinicians in practice remains a question for further real-world validation and clinical testing.
Citazioni salienti
Foresight is not meant to let patients self-diagnose or predict their future, but could help clinicians ensure a diagnosis is not missed or support continual real-time risk monitoring.— Zeljko Kraljevic, Research Fellow in Health Informatics, King's College London
Vuoi la storia completa?
Leggi l'originale su HospiMedica ↗
Health & Medicine