Re-imagining information and outcomes in healthcare
This blog post isn't a claim or a prediction about the future. It only serves to outline a 'broad-strokes' framework; I hope it stimulates your own thoughts on what the future of healthcare data may look like, and what we can strive towards.
Setting the scene
Healthcare systems exist to improve the health of their patients. What's not obvious is how good health can be effectively measured. To add a layer of complexity, all patients are biologically different, so health isn't a fixed outcome across a population. This makes measuring meaningful outcomes in healthcare quite challenging.
The framework: vertical and horizontal information
Let's take a typical path in a healthcare system, with a patient visiting multiple different service providers over time. At each interaction, data is collected and stored for future reference in the service provider's electronic health record (EHR). Figure 1 shows how these "transfer of information" events occur where the patient path intersects with the specific service provider.
Fig 1: Patients’ paths (P1, P2, P3) as they interact with service providers (SP1, SP2, SP3) over 2 consecutive time periods. In this case, patient 1 only interacts with 1 service provider over both time periods, patient 2 only interacts with service providers 2 and 3, and patient 3 initially interacts with service provider 1, but at the next interval interacts with service providers 2 and 3. Transfer of relevant information between service providers in patient 3’s case is not guaranteed.
In Figure 1, the 'horizontal' longitudinal patient information is fragmented across multiple sites. On the other hand, the 'vertical' service providers have full cohorts of information on their EHRs, but only intermittent snapshots for each individual over time. Lose lose. The problems with healthcare data fragmentation have been discussed extensively by many others, so I won't go into them here. It would suffice to say that transferring information from one service provider to another is high-friction and frequently entropic; information can be lost or corrupted and often the wrong amount of information is transferred (e.g. scans transferred without reports).
Outcome measures in healthcare
How does this 'horizontal-vertical' framework influence our assessment of how effectively patients are served by their healthcare system? Well, the outcomes we measure are a function of the information we collect. Given that the bulk of information is currently collected by the 'verticals' it's no surprise that our healthcare outcomes are commonly measured 'vertically' as well. For example, in clinical research, longitudinal outcomes are often estimated by the average change of the 'vertical' cohort over time, rather than by the 'horizontal' individual. This approach has brought into question the true effectiveness of a newly approved Alzheimer's disease drug, due to heterogeneity at the individual-level. Similarly in the hospital setting, audits and workflows are usually implemented to improve 'vertical' departmental outcomes, and direct measures of patient-centred outcomes are often relegated to flimsy, qualitative surveys. The problem is more pernicious than we think; even when developing value propositions for new technologies (including software and AI as a medical device), focus on workflow improvements and cost-effectiveness inadvertently take centre-stage as these are easier to measure. Apart from the obvious metrics (e.g. morbidity, mortality and diagnostic accuracy), patient-centred benefits are often inferred indirectly.
Refocusing on the right outcome measures
People are working hard to solve this issue. On one side, organisations such as the International Consortium for Health Outcome Measurement (ICHOM) aim to re-focus our attention on the most valuable patient-centred outcomes. However, the fact that these measures rely on vertical information means that extracting such outcomes remains difficult, indirect or overly simplistic. On the flip side, others are pushing to increase the 'inter-vertical' interoperability between EHRs. Whilst this is laudable, it doesn't necessarily address the issue of putting patient-centred outcomes first; merging vertical data streams has direct benefits for making vertical outcomes easier to measure, but only indirect benefits to horizontal outcome measures. Even if interoperability between service providers becomes seamless, the change in focus from population-level to individual-level outcome measures needs to be deliberate. This brings us onto a vision for how these two approaches could be bolstered by a change in perspective. It's time to think horizontally.
Individual-level outcomes are distinct from population-level outcomes
This conversation between Stephen Wolfram and Nassim Nicholas Taleb illustrates the reasons we need to focus on individual-level data if we want to improve the way medical science informs clinical practice and outcome measurement. For about 45 minutes from the timestamp, they discuss the relationship (or lack of) between scientific theories, medical 'science'/'statistics' and clinical practice. You can get the gist from the first 5-10 minutes, but if you have time, I can't recommend enough that you listen to the whole segment.
Taleb's main point is that statistical claims made on aggregate populations do not necessarily hold for the individual. Obviously, this isn't to say that we should never study aggregates - many biological variables follow well-described distributions (e.g. gaussians), and it helps to know where an individual patient may sit on the distribution. His point is that if the population-level distribution is the basis of your metric, you may lose some deeper, contextual information that would alter how effective a particular intervention is for an individual. Therefore, taking an individual-level perspective to measure outcomes also requires a change in how we treat data; we need to familiarise ourselves with concepts such as time-dependent averages and ergodicity. These approaches are slowly making their way into healthcare, ranging from intensive care to predicting long-term cardiovascular outcomes, but their application remains limited due to how the data are currently being collected (vertically).
Collecting and processing individual-level data - considerations for the future
Measuring individualised patient outcomes provides an opportunity to make a flexible healthcare system that works better for everyone. By putting the individual first, service providers can also increase the value (defined as outcome per unit cost) they offer to patients by focusing on bespoke outcomes. So why haven't we done this already?
Firstly, only service providers are currently able to offer the infrastructure needed to be trustworthy custodians of a population's health data (EHR software and secure data storage is expensive, and the amount of data to store is vast). Secondly, the individual has no means of collecting or storing their own information, let alone harmonise it with EHR data. Because of these two main barriers, the information needed to properly prioritise the individual doesn't really exist yet! However, as wearable/mobile technology, repeated testing and more detailed EHRs become commonplace, we are beginning to collect individual data that is rich enough to build better, personalised predictions and outcome measures.
A quick thought experiment
Taking these thoughts to an imaginary endpoint, we could envision a (not so) distant future where individuals become the custodians of their own health data, with built-in individualised predictive algorithms and outcome measurements.
Fig 2: A vision for the future. The same information is displayed as in Fig 1, but now with a horizontal focus. ‘Ongoing individual data collection’ represents the opportunity to collect from mobile applications or wearable devices, which can be combined with information gathered by service providers.
In this scenario, when the patient interacts with the service provider, the "transfer of information" event now happens in the orthogonal direction (Figure 2) compared to Figure 1. Service providers add information to the patient-owned data, not just onto a centralised EHR. Individualised outcomes and predictions can then be combined with the patient's choices and values to make the best clinical decisions. Not to be blinded by utopia, this thought experiment raises many tough questions: how would we validate individualised AI in supporting clinical decisions? What are the legal ramifications in terms of consent and data-sharing? What are the clinical risks and security implications? These frameworks don't yet exist because the technology isn't here yet, but it is something to think about. A more likely development in the foreseeable future may be a hybrid approach with an increased focus on assimilating horizontal information on the pre-existing vertical structure.
Wrap up
'Personalised medicine' has been the bedrock of the doctor-patient relationship ever since doctors and patients existed. 'Personalised medicine' is accelerating in clinical interventions (e.g. in genomics, pharmaceuticals, surgical implants and synthetic organoids). Now it's time to think about 'personalised medicine' in terms of the data we use to make clinical decisions and measure outcomes.
Hardian Health is clinical digital consultancy focussed on leveraging technology into healthcare markets through clinical strategy, scientific validation, regulation, health economics and intellectual property.
