Identifying health care-related harm: from patient record reviews to artificial intelligence

Hospitals always monitor the care they provide – keeping a close eye on any failures in care and creating action plans to reduce the chances of them happening again. Dr Helen Hogan and Chris Sherlaw-Johnson take a closer look at how efficient current processes are – and argue that advances in technology could soon see the UK make up for lost time.

Blog post

Published: 24/03/2020

Any patient in hospital will hope to receive the best care available. Of course, this will not always be perfect and in many cases that will not cause huge problems for the patient, but there will be occasions where wrong actions or lack of action can lead to severe harm, and even death.

NHS hospitals are constantly monitoring the care they provide, identifying problems and creating action plans to reduce the chances of them reoccurring. But how efficient are these processes? And how can advances in hospital information systems help?

Important yet imperfect

Information on harms may be recorded in hospital incident reporting systems or picked up from complaints. Some can also be gathered from hospital administrative data. However, it is individual patient case notes themselves that can uncover the broadest range of harms, particularly those related to omissions in care processes, which is much more common than harms caused by wrong actions. And there has been a long tradition of reviewing case notes to shed light on the quality and safety of health care services in general.

But it is not perfect. Case notes provide a rich picture of the technical care patients receive, but contain less detail on interpersonal care or the context in which the care is provided. Reviewers will tend to look more closely for errors where the outcome is judged as poor (for example, death), and disagreements between reviewers on both the nature of harm found and its preventability are common. Senior clinician time is also an expensive resource and case record review can be time-consuming, especially for long and complex admissions.

This inevitably limits the number of records that can be feasibly reviewed and the yield of identified harms. Paper records can also be illegible, incomplete or disorganised, with the potential consequence that organisations with higher quality records might erroneously appear to record higher levels of harm.

Some of these problems could be met by using routine hospital data to screen case notes for review – identifying patients who are more likely to have been harmed. These could be patients with unusually long lengths of stay, for example, or who are readmitted soon after discharge. There are some advantages in using such outcomes, as they focus on non-clinical measures of patient experiences that are less reliant on the quality of clinical coding.

Naturally, some of these patients may not have experienced harm – a long stay may reflect the difficulty of discharging someone who needs to go into care, or a quick readmission may only be a routine check that is mistakenly coded as unplanned. To filter cases still further, selection of readmissions could be restricted to people admitted with conditions that are more likely to reflect a previous failure of care.

In our recent study, however, we found that even with quite stringent selection criteria, the number of patients to review would still be very large in relation to the available resource. So, in practice, it may be sensible for a hospital to set a target number of, say, weekly reviews in advance and then randomly sample from those who meet the chosen criteria.

Future solutions?

In time, advances in information systems and technology may have considerable influence.

First, the wider adoption of electronic health care records should make it easier to identify risks and harm in patient records, and provide a clearer steer on which cases would benefit from in-depth scrutiny. Such systems can allow automatic real-time capture of clinical processes, such as the administering of medicines or physiological observations, and indicate failures in timely response and intervention. Structured information about clinical care along treatment pathways will allow clinicians to identify, at the touch of a button, whether patient management varies from evidence-based criteria.

Learning from the circumstances that lead to such variation, as well as the impact on patient outcomes, can support the modification of guidelines and improved adherence. This approach becomes even more powerful when information on staffing levels or skill distribution at the time of the variation in care is also available – providing insight into the context in which care is taking place and its impact on clinician behaviours.

Second, rapidly advancing artificial intelligence and natural language processing may finally provide an efficient way to search the non-structured, free-text parts of case records for evidence of harm, and for the first time provide a picture of its real scale and scope.

In the last two decades, these approaches have been developing in the USA. Given the challenges of mining such information, not least the need to unravel the multiple clinical terms and abbreviations used for similar conditions, such work has often, inevitably, kept a narrow focus. Typical applications include identifying medication errors or post-operative complications such as infections.

Rapidly catching up

The UK has lagged behind, but is likely to become a rapid adopter now that technological transformation has been placed at the heart of the Long Term Plan, capturing ambitions for the service over the next decade. Programmes such as the Global Digital or Local Health and Care Record Exemplars will lead the way in adopting and spreading world-class digital technology, interoperability, and new possibilities for securely sharing data across organisations.

With the right algorithms combining structured information in case records with unstructured narrative, there is scope for considerable progress in the early identification of patients who are at risk of harm, leading to earlier intervention to prevent it. The richness of the information will also allow a better understanding of which factors create safety risks, and reveal clearer targets for quality improvement interventions and ongoing system monitoring.

Dr Helen Hogan is an Associate Professor in Public Health at The London School of Hygiene & Tropical Medicine, and Chris Sherlaw-Johnson is a Senior Fellow at the Nuffield Trust.

Suggested citation

Hogan H and Sherlaw-Johnson C (2020) “Identifying health care-related harm: from patient record reviews to artificial intelligence”, Nuffield Trust comment.