Patient stratification

Stratification is needed when a population is heterogeneous across several dimensions at once and classifying it by a single variable, whether age, a diagnosis or a score, does not capture the real profiles that coexist within it. It applies when you want to design differentiated responses for groups of patients with similar care needs, rather than applying a uniform protocol to a population that is not uniform. Classifying by one criterion groups together patients who are alike on that variable but differ on everything else. Patient stratification combines clinical variables, comorbidities and resource use to identify internally homogeneous subgroups. Treating heterogeneous populations with overly uniform protocols can waste resources and leave specific needs less well covered. Stratification provides a quantitative basis for prioritising effort by clinical profile, complexity or resource use.

You need clinical, comorbidity and resource-use variables for the population you want to segment, complete enough and with enough coverage for the resulting profiles to be interpretable rather than a reflection of recording patterns or missing data. The client provides these data; the service selects and codes the variables, segments the population, and validates and interprets the profiles. Selecting and coding the variables that feed the model is a particularly sensitive step: the result of a segmentation depends more on which variables go in and how they are represented than on the specific clustering technique. We also check whether the available variables give balanced weight to the relevant clinical, social and resource-use dimensions, because a segmentation can only find the profiles the data allow it to observe. The database does not need to arrive clean, but it helps to be clear about which dimensions of the population should shape the profiles, because that decision conditions the whole result.

There is no objectively correct number of profiles: the number of groups is an informed methodological decision, not something the data reveal on their own. This applies to any segmentation based on unsupervised clustering, where, unlike a predictive model with a known outcome, there is no external answer to validate against. Choosing the number of groups combines statistical criteria of cohesion and separation with a requirement that carries equal weight: the resulting profiles must make clinical sense and be distinguishable in practice. The decision can draw on internal validity indices, the stability of alternative solutions and a clinical review of how interpretable each profile is. A large number of clusters may fit the data better and be useless in the clinic; too few can dilute clinically relevant differences. The aim is not the number that optimises an index, but the one that produces profiles the clinical team recognises and can manage differently.

A profile is more credible when it holds up under reasonable changes to the data, the parameters or the validation period, not when it appears once in a single analysis. This applies to all patient stratification, because clustering techniques always return groups, even when there is no real structure to segment: the algorithm partitions the population even if it is homogeneous. That is why validation is the core of the work, not an add-on. Cluster stability is assessed through sensitivity analysis, checking whether the profiles persist when parameters or subsamples change, and external validation is sought in independent cohorts or different time periods. A profile that replicates in new data is a finding. One that appears only in the dataset where it was discovered should be interpreted with caution, because it may depend on the method, the sample or the coding of variables. Stability and reproducibility are what separate a useful typology from an arbitrary partition.

A profile is clinically actionable when, as well as being statistically sound, it has a clear clinical interpretation and supports a care decision that differs from the one for other profiles. It applies as a quality criterion for any segmentation intended for clinical use: a mathematically correct cluster with no clinical reading cannot inform any decision. That is why each profile is interpreted in clinical terms, looking at what characterises those patients and how they differ from the rest, rather than delivered as a numeric label. The difference is the one between "group 3" and "multimorbid patients with high resource use and social frailty", which can guide a decision on follow-up, prioritisation or resource allocation. A profile should not become a fixed label on the patient: it describes a useful pattern in a population and should be reviewed if the data, the inclusion criteria or the care context change. The service identifies, validates and interprets the profiles; the specific clinical response for each one is defined by the clinical team, who know the context and the resources.

No. The service identifies and validates the profiles, interprets them clinically and provides the evidence for the care team to design its care pathways or differentiated protocols; designing the care pathway is the hospital's responsibility. The distinction is deliberate: defining a care pathway requires knowing the centre's circuits, resources and constraints, and that knowledge belongs to the clinical team, not to the analyst. What the service contributes is the foundation that differentiation is built on: which profiles exist, what characterises them and, where relevant, an analysis of pathway adherence and the factors associated with deviation. This separation protects the client, because a pathway designed without the operational knowledge of the department is rarely workable. The service delivers a map of profiles that is methodologically validated and clinically interpreted; the care pathway is designed and implemented by those who know the ground.

No. Patient stratification models group a population into profiles or risk levels to organise management and resource allocation: they answer "what profiles or levels of management exist in this population?". Clinical risk and readmission prediction estimates each patient's probability of a specific future event, such as deterioration, readmission or death, within a defined time window: it answers "how likely is this patient to experience this event?". They do not answer the same question. Stratification organises a population into interpretable profiles that are useful for planning care or resources; prediction quantifies an individual risk. The two can be combined, since a risk prediction can feed a stratification and a stratum can trigger a specific predictive model, but one orders the population by management profile and the other anticipates an event in the individual.

Patient stratification is designed from the outset around how it fits into the centre's clinical workflow and information systems, not as an analysis that ends in a document. This applies when the aim is for the profiles to be used in practice, whether in the clinic, in resource allocation or in prioritisation, which is where segmentation adds real value. To that end, the profiles are defined so that the team can recognise and apply them, and the model is documented so that it can be re-evaluated as the data are updated, so that it does not go out of date unnoticed when the population changes. The service delivers validated, interpretable and reproducible profiles, ready to be integrated into the system if the centre has the technical and operational channels to do so; the management decision for each profile and its use in the clinic belong to the clinical team. Stratification has to work in the clinic, not end up in a report.