Segmentation is treated as analytics. It is, in fact, a decision about how the operating model carries risk. A working-paper position.
Segmentation is treated as analytics. It is, in fact, a decision about how the operating model carries risk. A working-paper position.
The familiar failure
Most population segmentation projects end the same way: a handsome cluster analysis, a slide with named segments, a leadership team that nods, and an operating model that carries on exactly as before. The analysis was real. The segments were statistically defensible. And nothing changed, because the organisation treated segmentation as a data exercise when it is actually one of the most consequential design decisions a health system can make: the decision about which groups of people the operating model is built around, and therefore where clinical and financial risk is carried, by whom, and on what terms.
Our position is simple. If a segmentation does not change who does what, for whom, with what team, under what funding logic, it was not segmentation. It was decoration.
What segmentation actually claims
The intellectual foundation is older and sturdier than the current analytics fashion. The Bridges to Health model divided a population into eight segments, from people in good health through stable chronic illness to long-term frailty and failing health near death, and made the essential observation that each segment has its own definition of optimal health and its own priorities among services (Lynn et al., 2007). That sentence carries the entire operating-model consequence. If what "good" means differs by segment, then a single care model, a single workforce configuration and a single funding logic cannot be optimal for all of them. Segmentation is the admission that the average patient does not exist, and the commitment to stop designing for them.
Notice what this implies about the direction of the work. The segments are not discovered in the data and then admired. They are chosen as units of design: for each, the outcome that matters, the care model that produces it, the team that owns it, and the economics that sustain it.
Why risk scores are not segments
The most common substitute for segmentation is risk stratification: rank the population by predicted cost or admission risk, draw a line at the top 5%, and build a programme for the "high-risk" group. Two findings from the evidence explain why this keeps disappointing.
First, the high-risk population is not a segment; it is a statistical residue. When researchers took the top 5% of a risk-stratified English population and analysed how those people actually used care, they found four distinct groups with different usage patterns across emergency, elective, outpatient and primary care, patterns stable enough over time to design services around (Vuik, Mayer and Darzi, 2016). A single high-risk programme built for that population is built for nobody in it. The clinical need, the behavioural reality and the right intervention differ across the four, even though the risk score that grouped them was identical.
Second, risk is not amenability. Predictive models identify people likely to generate cost; they say nothing about whether preventive intervention will work for them. The corrective concept, impactibility, asks which of the at-risk are actually amenable to the programme on offer, and the analysis of it carries an uncomfortable warning: filtering for impactibility can quietly deprioritise exactly the patients with the most complex needs, so the choice of filter is an equity decision, not a technical one (Lewis, 2010). An operating model that targets "the impactible" without saying so out loud has made a values choice in a spreadsheet.
Both findings point the same way. The analytics can find the groups. Only the organisation can decide what it owes each of them, and that decision is the operating model.
The operating-model decision, made explicit
Taken seriously, a segmentation forces five commitments per segment, and a health system's leadership should be able to recite them for every segment it claims to serve.
The outcome that defines success. For the healthy segment, staying healthy at near-zero friction; for stable chronic disease, avoided deterioration and undisrupted life; for frailty, function, dignity and the avoidance of harmful escalation; for the person near death, comfort and control. These are different products. Pretending otherwise is how systems end up measuring everything by throughput.
The care model that produces it. Episodic acute medicine for the segment that needs episodes; continuous, relationship-based management for the segments whose risk lives between visits. The encounter remains the unit the model is designed around; what changes by segment is the cadence, the setting and the team composition that protect it.
The team that owns it. Ownership means a named service with authority over the pathway, not a coordination committee spanning silos that all keep their own incentives.
The economics that sustain it. This is the decision most segmentations never reach: where risk is carried. Episode-based funding fits episodic segments; it structurally punishes whoever invests in keeping the chronic and frail segments out of hospital, because the saved admission is someone else's saved cost. A segment whose value is created by prevention needs a funding instrument that pays for prevention, whether capitation, bundles across a period, or explicit contracts for avoided utilisation. Choosing segments without re-choosing funding instruments is deciding, silently, that the operating model will fight its own strategy.
The measurement that keeps it honest. Per-segment outcomes and per-segment economics, reviewed as seriously as the aggregate P&L, because the aggregate is precisely what segmentation exists to see through.
The sequencing discipline
The order of operations matters as much as the content, and it mirrors the sequencing we argue for everywhere: design the care model first, then build the machinery around it.
Segment on need and behaviour, using the data to test and refine human-legible segments rather than to generate statistical clusters no clinician recognises. Define the outcome per segment. Design the care model and team per segment. Align the funding and risk-bearing per segment, negotiating with payers segment by segment rather than tariff line by tariff line. Only then choose the technology, the analytics and the monitoring that the segment's care model needs. A system that runs this sequence backwards, buying the analytics first, gets what most systems have: a risk score in search of an operating model.
There is also a duty of honesty about data limits. Segmentation models are built on utilisation history, and utilisation reflects access as much as need; the person who never reaches care generates no data and appears low-risk. A segmentation that only sees the demand that already arrived will optimise the system around exactly the people it already serves. This is the same equity trap Lewis identified in impactibility filtering, operating one level earlier (Lewis, 2010).
The test
We hold segmentation work to one test, and we invite boards to apply it to any segmentation they have paid for. Pick a segment. Ask what is operationally different for the people in it: which team, which access route, which cadence of contact, which outcome on the dashboard, which funding instrument. If the answers are the same as for every other segment, the organisation does not have segments. It has labels.
Segmentation done as an operating-model decision is among the highest-leverage moves available to a system under the pressures every system now faces, precisely because it replaces the average patient, who does not exist, with real groups whose needs can actually be met. That is not a data exercise. It is the structure of the promise a health system makes, segment by segment, to the population it serves.
References
- Lewis, G.H. (2010) '"Impactibility models": identifying the subgroup of high-risk patients most amenable to hospital-avoidance programs', The Milbank Quarterly, 88(2), pp. 240-255. https://doi.org/10.1111/j.1468-0009.2010.00597.x
- Lynn, J., Straube, B.M., Bell, K.M., Jencks, S.F. and Kambic, R.T. (2007) 'Using population segmentation to provide better health care for all: the "Bridges to Health" model', The Milbank Quarterly, 85(2), pp. 185-208. https://doi.org/10.1111/j.1468-0009.2007.00483.x
- Vuik, S.I., Mayer, E. and Darzi, A. (2016) 'Enhancing risk stratification for use in integrated care: a cluster analysis of high-risk patients in a retrospective cohort study', BMJ Open, 6(12), e012903. https://doi.org/10.1136/bmjopen-2016-012903