Blogs

The Solution for the AI Challenge must be a catalyst for a new theoretical framework. Risk stratification alone will not afford options for actions that will achieve the desired outcomes. The CMS data is factual but not curated. Exploratory analysis is best done before machine learning is set loose on the data; this is true of any statistical analysis as well.

Branches of AI that include model building are causal paths and counterfactual analysis which are transfactual; The CMS beneficiary data is transactional and includes some metadata useful for analysis, but cannot be relied on as the only component. Including metadata that is real but not yet known will be a critical component of an effective solution. For example, actions or services that are driven by mechanisms other than medical, as reflected in claims codes, include: psychosocial effects, individual preferences of patients and providers, unexpected events that occur that are not amenable to prediction (e.g. auto accidents), noncompliance with treatment recommendations, etc… the list is endless.

Curation reserves a place for the factual and the transfactual. This is what Health Care in 3 Dimensions (HCn3D) is intended to do. Within each dimension curation can be named structures or groupings. Metadata, which adds dimensions in the form of the patient journey can be grouped by disease, cost corridors, and georeferenced (e.g. by county). This first dimension is patient centered, and is the place to curate the “n of 1”, a necessary concept to design effective solutions that reflect the uniqueness of the individual patient. It is not amenable to population averages. The AI community will accept the n of 1 unwillingly, but this will resonate with family physicians. This is the essence of the Challenge; can AI be a “big tent” that accepts all resolutions and scales of data including the level that is meaningful to the patient and the physician, but yet is merely anecdotal for the BIG Data only community. Acceptance by physicians and patients will require the output of AI to be not only explainable, but valuable as well.

The next dimension is the network, which is provider centered, defined by scale as local or by characteristics of the providers. In having only general descriptors of providers such as specialty, group practice, benchmarks of cost (as in BPCI), etc… the characteristics of providers that are unique are lost, and is similar to the problem of the “n of 1” for the patient. Solutions will be more effective if the specificity of individual providers is treated as transfactual, in other words, data that is not known from claims sets, but determined by machine learning techniques, will show for similar providers that outcomes will differ. Curation here will group the providers into virtual networks where the individual characteristics will be the attributes of the network group. This is another way of saying that provider networks can be more usefully analyzed within a meaningfully grouped hierarchical structure. Note, this second dimension, of the network of providers and the characteristics they employ in patient decisions, is local and therefore considered “small data”.

The 3rd dimension, the scale of large populations of CMS beneficiaries, is the place for techniques of BIG Data analysis. But, is this the place to find solutions? There are many opportunities to use this data, but as an enabler of meaningful metadata to benefit provider networks functioning in unique geographic sections with unique population demographics, and with many sources of practice and outcome variations; the AI solutions for what are transactional issues will only be possible by including all dimensions.

CMS's intent to use AI to accomplish Challenge objectives, faces unrealized and unapparent obstacles of both small and large data sets in the healthcare commons. AI is a “big tent” that uses techniques appropriate for large scale populations as well as tools that target small populations and even single causal agents. To benefit from all scales in the healthcare commons, from CMS populations, to local networks, and the patient (n of 1); the AI tools of Health Care in 3 Dimensions can provide the scalability needed to bridge the entire commons. As recognized by the AI Health Outcomes Challenge, “medical decision-making and population health management have not significantly changed”. Clearly more can be done with the massive data sets at hand, but for the data to be useful, a systematic method of curation must be employed that understands the multiple levels of the healthcare network from the entire population of CMS down to the patient (n of 1), including the transactions that occur along the patient journey.

To demonstrate the need for a method of curation, in assessing payment and delivery methods, the example of reducing unplanned admissions is a good space in which to start. The diversity and complexity of the healthcare commons is well exemplified by the problem of unplanned admissions and is a good surrogate for the thousands of potential targets of opportunity for cost reduction.

Health Care in 3 Dimensions provides the framework to apply theory to practical problems such as reduction in unplanned admissions. This approach will resonate with family physicians and can be tailored to them as the target audience for AI.