As I age I start to worry about dementia. Yes, I forget names – but generally if I wait a few minutes they will come back. But I am concerned for the Health System – it appears to be suffering from cognitive decline. It is now routine for clients to be seen by clinicians who have never seen them before – the “anonymous consultation”. With the loss of traditional General Practice, ubiquitous but poor quality electronic medical record systems and increasing staff turnover everywhere, continuity and client relationships have suffered. Past history is “forgotten” – diagnosis relies on the presenting symptoms and observations. Clinicians are less sophisticated than they were – they are not likely to consider alternative diagnoses such as masquerades and unusual presentations. Multimorbidity is less well managed by the “anonymous” clinicians as they struggle to learn the detail of a complex client. Expertise and the knowledge of and a relationship with a client has been replaced by complex rituals of checkups and Careplans, items and counterchecks. But important issues still seem to fall through the cracks – it appears these systems are not a safety net. Lumps that should have been investigated and removed are forgotten for years before they are finally dealt with. Or perhaps they are never confronted and the patient dies – even then the potential cause is not identified. Clients may present repeatedly with the same issue with the system apparently not remembering previous investigations and outcomes.
At a policy level the constant loss of corporate knowledge due to turnover means that the same issues come up again and again – they appear to be immutable – the previous solutions and outcomes have been forgotten.
Does Past History Matter?
We spent a lot of time in our training learning a systematic approach to Medical History. A major part of this was obtaining past history. It was regarded as an important predictor of the likely diagnosis on the presenting occasion and there were often other issues that required followup. For a complex multimorbid patient it was important to obtain a full picture of the clients medical history to plan and prioritize interventions. But it appears from the behaviour of many clinicians that this is no longer regarded as important. Past history is almost routinely ignored and the client treated for their presenting complaint. The system relies on programmed interventions such as Careplans to manage ongoing issues. But these are far from perfect – in practice they cannot be easily tailored to an individual and in many cases relevant past history is missed and forgotten. Clinicians are encouraged not to think independently but to just follow the prompts. In a previous article I discussed Complexity in Medicine – if indeed many clients can be regarded as “Complex” then a programmed approach is likely to fail. Good management of Complex Multimorbid clients (the majority of those over 40) requires a map of their issues and a tailored “thinking” approach at every encounter. When serious lifethreatening illness presents, it often does so over several clinic visits over several days. When mistakes occur in this situation it is usually due to different clinicians not referring to the events on previous days – again, Past History Matters.
What is AI ?
AI (Artificial Intelligence) is the buzzword of the times. But many people dont have a clear understanding of what it means. There have been massive advances in this field in the last 20 years and it continues to rapidly evolve. Essentially it involves several large datasets. A suitable dataset is separated into categories with the desired outcome. This is then used to “train” a mathematical algorithm as a “black box”. Once the model has “learned” the pattern to an acceptable degree, it is supplied with “wild” data to achieve the same outcome as the training dataset. These AI algorithms involve many simultaneous multidimensional calculations similar to 3D games – hence the use of video chips optimized for this. The American company Nvidia has seen its shareprice skyrocket – it started life as a video card and chip maker. More recently there has been a proliferation of LLMs (Large Language Models). These incorporate large numbers of language concepts and the relationships between them. They form the basis of smart chatbots and assistants for tasks as diverse as programming to marketing. These can ingest large amounts of data such as text and extract relevant concepts from them.
The search problem for clinicians
If we accept that Past History does matter, then the first task after the elucidating the presenting complaint from a client is to obtain a past history. The clinician encountering a complex multimorbid patient for the first time must try to get a complete Medical picture of their client. They are searching for “anything” relevant. This involves interrogating the patient and the medical record for important concepts . These are found in letters, pathology and imaging results, progress notes and documents. If the clinician is lucky, a previous expert clinician has created a summary of the relevant issues. The search may involve working across several record systems and interfaces. But this remains difficult due to silos and restricted access due to security concerns. The situation has not improved in recent years – major development projects have not allocated resources to interfaces and vendors remain resistant to interoperability for commercial reasons. In general the record has been designed by bureaucrats searching for specific data – they have a different search problem to the clinician. Data is hidden in easily searchable “items” rather than free text (which is frowned upon). There is poor interface design, with lots of irrelevant headings, poor formatting, poor labelling, administrative entries and many “null” values. A clinician has to search this very “noisy” environment for relevant data.
How do we record a Clinical encounter?
It is important to record what happens in a clinical interaction for various reasons. Perhaps the most important is Clinical, to assist with future clinical interactions. There is also a legal imperative in case the encounter should be contested in future. Here, too much detail is never enough and this drives much of what is in the record. Third, the record is used for administrative and research purposes. Again this drives much of the content of the record even though it’s primary purpose is ostensibly clinical. How to record all this? The ultimate is to take video of a consultation. This is not performed often for many reasons including consent, storage and clinician resistance. In most cases a written account is entered by the clinician into an electronic record, with details of demographics, items describing various actions and measurements and free text “Progress notes” describing the interaction. Finally, most systems enter a “reason for encounter’ which may be multiple. Here codesets such as ICD or ICPC2 form the basis of a picklist. So an encounter can be described in various ways, from a full video to a single code.
Possible Use Cases for AI
(1) A “Past History Engine”
Could an AI generated Past History summary help the Clinician searching the record of a “Complex” patient? Could it prompt the less sophisticated Clinician with relevant information?
Large Language Model (LLM) based systems are now being widely used in various areas of commerce and legal practice to extract relevant concepts from large datasets of text. It would be possible for such a system to ingest an entire medical record with associated documents and even the documents linked to other systems. It does so much more efficiently than a human – in fact it may be impossible for a human clinician to perform this task efficiently in a large record with hundreds of entries and thousands of data items.
Territory Kidney Care
This system was set up to provide a summary of the past history of the many clients in the Northern Territory of Australia (NT) who were suffering from or at risk of renal disease. The aim was to help clinicians in managing them and to provide an early warning of deterioration by automated prompts. Some 10,000 clients were entered onto the system. Data on these clients was obtained from various sources with the relevant consents of their treating organizations. Data included problem codes, clinical measurements, pathology results and Medicare billing – this was entered into a single database which in turn was interrogated to provide a summary and timeline of various parameters available via a web interface. This effort was privately funded and mentored by a well known research organization. It cost a small fraction of typical comparable record systems (approx 1% of the cost the NT Govt new system Acacia!). The web interface was designed by a Darwin company – it was clean, simple and easy to use. While this was not a full Electronic Record System, it did show what was possible at minimal cost and with good interface design. By contrast, Government Health IT systems are typically expensive and difficult for users to navigate.
This system did not use AI – just basic data algorithms and clean interface design with minimal administrative “noise”. It also overcame the apparently intractable problem of getting data across interfaces between different systems – data was obtained from most of the large Health organizations in the NT. It gave a useful insight into where clients were attending.
If such an approach was broadened into large record systems and an algorithm based on an LLM to mine all the text into the record, we could obtain a good summary of a client’s past history. This would improve clinical management, efficiency and safety. It could go a long way towards providing the continuity that has been lost in modern Health systems
(2) “Recall Engine”
One of the mechanisms used in records to maintain continuity is the “recall”. It is entered for a specific date in the future with details of the clinician targeted and the actions to take. But current systems generate large numbers of often poorly targeted or incorrect recalls. The recall may be “serviced” but not removed or become unnecessary. As a result of this large (even overwhelming) number of recalls the system is difficult to use and is often ignored altogether. In my own surveys of records, outstanding recalls are not serviced in the majority of encounters. Could a smart or AI enabled system improve this situation? It could be based on the “Past History Engine” above and generate a minimum of recalls that are relevant and targeted. It could also generate a single optimized “Careplan” based on their known problems for the client with relevant planned interventions.
(3) Location
In a large system of clinics such as NT Government Remote Health, a particular client has a “usual clinic” which is tasked with delivering scheduled interventions such as Careplans . But many people visit several clinics, often in different health systems across borders and with different provider organizations. Some are “transient” with no identified clinic “owning” them. These tend to be a high risk and high needs group. One approach to service clients better would be to adopt an opportunistic approach to care and deliver all interventions wherever they present. But if we are to persist with the “programmed” approach we need to identify which clinic the client should be attached to – all clients, transient or otherwise, should belong to a nominated clinic. An AI algorithm looking at time series location data could possibly predict where the person is next likely to attend and target relevant interventions there.
Conclusion
The Health System is suffering from cognitive decline driven by increasing staff turnover and administrative complexity. Electronic Health Record (EHR) systems are not up to the task of replacing a long term relationship with an effective means of maintaining continuity. This decline could be reversed with changes in policy and business rules, smart EHR design and targeted use of AI to improve data visibility and prompt clinicians appropriately.
TjilpiDoc 