Related Experiment Videos
Measuring large language model uncertainty in women's health using semantic entropy and perplexity: a comparative
Jahan C Penny-Dimri1, Magdalena Bachmann2, William R Cooke1
1Oxford Digital Health Labs, Nuffield Department of Women's and Reproductive Health, University of Oxford, Oxford, UK.
The Lancet. Obstetrics, Gynaecology, & Women'S Health
|July 6, 2026
Summary
Semantic entropy effectively detects inaccuracies in large language model outputs for obstetrics and gynaecology, enhancing patient safety in clinical AI applications.
Area of Science:
- Artificial Intelligence in Medicine
- Natural Language Processing
- Clinical Decision Support Systems
Background:
- Large language models (LLMs) show promise for clinical decision-making but risk generating unsafe, incorrect outputs (hallucinations).
- Existing uncertainty detection methods fail to identify meaning-level inconsistencies in LLM responses.
- Semantic entropy, a novel metric for meaning variation, requires validation in clinical settings.
Purpose of the Study:
- To compare semantic entropy with perplexity in detecting inaccuracies in LLM-generated answers for obstetrics and gynaecology.
- To evaluate the utility of semantic entropy as an uncertainty metric for clinical LLM applications.
Main Methods:
- Assessed semantic entropy using a dataset from UK Royal College of Obstetricians and Gynaecologists (MRCOG) exams.
- Generated responses using GPT-4o, computing semantic entropy from meaning-based clusters.
- Compared semantic entropy and perplexity using AUROC and accuracy, with expert clinical adjudication for a subset.
Main Results:
- Semantic entropy demonstrated superior discrimination of incorrect responses (AUROC 0.76) compared to perplexity (AUROC 0.62).
- Expert validation showed near-perfect discrimination for semantic entropy (0.97) versus chance-level for perplexity (0.57).
- Semantic entropy outperformed perplexity across various question types and response lengths; a simplified variant also performed well.
Conclusions:
- Semantic entropy provides a robust method for identifying uncertain or misleading LLM outputs in clinical settings.
- This metric can enhance AI-assisted clinical decision-making by flagging unreliable responses for human oversight.
- Semantic entropy serves as a practical safeguard for deploying LLMs in medicine, establishing a framework for future uncertainty estimation.
Related Concept Videos
01:10Propagation of Uncertainty from Systematic Error
The atomic mass of an element varies due to the relative ratio of its isotopes. A sample's relative proportion of oxygen isotopes influences its average atomic mass. For instance, if we were to measure the atomic mass of oxygen from a sample, the mass would be a weighted average of the isotopic masses of oxygen in that sample. Since a single sample is not likely to perfectly reflect the true atomic mass of oxygen for all the molecules of oxygen on Earth, the mass we obtain from this particular...
00:54Uncertainty: Confidence Intervals
The confidence interval is the range of values around the mean that contains the true mean. It is expressed as a probability percentage. The interpretation of a 95% confidence interval, for instance, is that the statistician is 95% confident that the true mean falls within the interval. The upper and lower limits of this range are known as confidence limits. The confidence limits for the true mean are estimated from the sample's mean, the standard deviation, and the statistical factor 't,' or...
02:46Uncertainty in Measurement: Reading Instruments
Counting is the type of measurement that is free from uncertainty, provided the number of objects being counted does not change during the process. Such measurements result in exact numbers. By counting the eggs in a carton, for instance, one can determine exactly how many eggs are there in the carton. Similarly, the numbers of defined quantities are also exact. For example, 1 foot is exactly 12 inches, 1 inch is exactly 2.54 centimeters, and 1 gram is exactly 0.001 kilograms. Quantities...
00:59Propagation of Uncertainty from Random Error
An experiment often consists of more than a single step. In this case, measurements at each step give rise to uncertainty. Because the measurements occur in successive steps, the uncertainty in one step necessarily contributes to that in the subsequent step. As we perform statistical analysis on these types of experiments, we must learn to account for the propagation of uncertainty from one step to the next. The propagation of uncertainty depends on the type of arithmetic operation performed on...
03:37Uncertainty in Measurement: Accuracy and Precision
Scientists typically make repeated measurements of a quantity to ensure the quality of their findings and to evaluate both the precision and the accuracy of their results. Measurements are said to be precise if they yield very similar results when repeated in the same manner. A measurement is considered accurate if it yields a result that is very close to the true or the accepted value. Precise values agree with each other; accurate values agree with a true value.
03:34Uncertainty in Measurement: Significant Figures
All the digits in a measurement, including the uncertain last digit, are called significant figures or significant digits. Note that zero may be a measured value; for example, if a scale that shows weight to the nearest pound reads “140,” then the 1 (hundreds), 4 (tens), and 0 (ones) are all significant (measured) values.