In many, and perhaps the majority, of cases, this is entirely appropriate as a first approximation. At some level, we’re always aware that the patient pool isn’t a monolithic slab of people. There’s a flux of some kind, patients moving in and out of the pool, and movement between the segments within the pool, but it’s usually sufficient to think of the population as being at some kind of equilibrium, the only significant driver being the drift of demographics. But, what about cases where the population isn’t at equilibrium? Or the products we’re modeling have the potential to alter the natural history of the disease? Enter Patient Flow models.
Simply defined, a patient flow model is a longitudinal model which looks at individual cohorts of incident patients, then tracks them over time as they move from one state to another, through their ultimate outcome. By summing all incident cohorts over a period of time, we can build up cross-sectional slices of patients that meet our particular segmentation criteria at a point in time (or averaged over a defined time period). What this does is allow us to model the way things that happen today affect the outcome tomorrow.
- Changes in disease incidence or mortality beyond simple demographics
- Progressivity of the disease state
- The potential for therapy to disrupt the natural history of the disease
- Significance of prior therapy history in later lines of treatment
Consider Chronic Hepatitis C in the US as an example. There was a huge spike in incidence during the Vietnam War, due to infected blood products being given to wounded soldiers, and another spike around 1990 due to the popularity of injection drug use at the time. Since then, incidence of new cases has decreased enormously. Given that the prevalence of HCV is calculated as Prior Prevalence + Incidence – Mortality – Cure, the long-term outlook for HCV prevalence is driven by the bolus of already-infected patients, and their cohort-specific demographics over time.
Some diseases are sufficiently stable that, upon diagnosis, itâ€™s difficult to determine how long the patient has had the disease, because of the lack of progression once infected. Chronic Kidney Disease, on the other hand, is characterized by a steady, near-linear in fact, decline in Glomerular Filtration Rate, and is staged accordingly, culminating in the need for dialysis and transplant. If your output metric of interest is the dialysis population, anything that changes along the pathway in CKD will affect the size of the dialysis pool, at some point in the future.
Consider HER2+ Breast Cancer. Before the launch of Herceptin, the outlook for even early-stage patients was discouraging. HER2-targeted therapies offer higher cure rates in local disease, and delayed progression in metastatic disease. Think for a moment about 2nd Line treatment in the metastatic setting. Historically, there has been a steady flow of progressing patients from the 1st Line market, but what happens the year Herceptin launches? As patients are treated in the 1st Line setting, those treated with Herceptin will experience longer remission, leading to a decrease in the short-term progression to 2nd Line treatment. Eventually, those patients will progress, and be added into the 2nd Line pool at a later date, but the effect that has on the 2nd Line pool depends on the penetration of Herceptin in the 1st Line market, and the Progression-Free Survival curves for Herceptin, relative to conventional treatment.
Returning to the HCV example, novel antiviral therapies have shown higher rates of Sustained Viral Response than existing therapy. The implication is that the choice of treatment in the 1st Line setting will affect the number of patients entering the pool of prior treatment-failures, as well as affecting the overall prevalence of the disease. The opportunity to re-treat patients is therefore contingent on the choice of prior treatment. On top of that, the number of patients eligible to be treated with Product X in 2nd Line depends not only on the number of patients treated with Product X in 1st Line, but also the relative difference in efficacy between Product X and other products.
Ultimately, what these considerations have in common is the notion that something is changing the disease dynamics to a potentially meaningful degree. The more that’s true, the less reliable a non-patient flow approach becomes.
Infectious Disease – Infectious diseases are often, but not exclusively candidates for patient flow approaches. HSV, for example, is not life-threatening, progressive, curable, or subject to wide variation in incidence. A conventional meta-analytic epidemiological approach is perfectly adequate in this case. HIV and HCV, on the other hand, are strong candidates for a patient flow approach. Incidence rates have varied widely over recent decades, and the evolution of treatment has extended life expectancies. Early in its history, HIV life expectancy was less than a decade. These days, it’s considered a chronic, but manageable condition. HCV has been discussed already. In both cases, treatment rates and choices today affect the pool of patients eligible for treatment tomorrow.
Cardiovascular/Metabolic Disease – Traditionally, there has been little consideration of patient flow in these patient pools. Within the rather broad group of ailments comprising the CV/Metabolic marketplace, there strongest candidate for patient flow is Type 2 Diabetes. Increasing incidence over the past two decades has caused the prevalence to increase rapidly, and the age distribution of the population to skew towards younger patients. Advances in treatment for T2D, as well as its life-threatening sequelae have impacted mortality rates accelerated this process. The increased activity in the market has led to a more robust treatment algorithm, and extended the path from diagnosis to insulin dependence. However, these changes are more gradual than we see in other markets, and treatment has not materially affected life expectancy since insulin was first made widely available, so the decision on whether to adopt a patient flow can be argued both ways, ultimately depending on the degree of detail you need on the patient demographics within the population.
Beyond T2D, the trends in the CV marketplace are gentle enough, from the standpoint of both patient characteristics and clinical outcomes, that unless the product you’re forecasting has a specific pattern of use that depends on knowing the longitudinal patient history, patient flow is unnecessary.
Other markets – To touch briefly on other major markets, Psychiatric indications are generally not candidates, although we have built patient flow simulators to model the way Schizophrenic patients move in and out of treatment, as a function of treatment persistence/compliance and the relative difference in relapse rates between compliant and non-compliant patients in the community setting, but this was largely done because the available secondary data is so old that they don’t reflect a marketplace where the better-tolerated atypical antipsychotics are widely used.
Autoimmune marketplaces are potentially strong candidates in cases where disease-modification is a feature of treatment. The rise of biologics has suggested a place for patient flow models within the treated pools, as a means to identifying the fraction of eligible patients over time. Rheumatoid Arthritis is the most high-profile example of this, but the idea extends elsewhere.
This is by no means a complete or definitive list of where a patient flow is appropriate, but it serves to demonstrate the wide, but not universal applicability of the approach.
Detail – there is no question that a cross-sectional analysis can answer that can’t be better answered by a patient flow model, and there are many questions that can only be answered by patient flow analysis.
Accuracy – in those cases where a patient flow model is used (and built properly), there are pitfalls that a conventional approach will miss entirely. In HCV, for example, increases in treatment rate today deplete the pool of available patients tomorrow. The impact of this can’t be predicted or quantified without explicitly modeling the process that leads to it, but it’s a significant driver of the market. Not using a patient flow when it’s needed can easily multiply your forecast error by an order of magnitude.
Flexibility – Once the flow has been modeled, the outputs you choose can be easily calculated in as much, or as little, detail as needed for the forecast. If the forecast segments change, developing new outputs that remain completely consistent with the prior segmentation is very easy. Doing this with a conventional analysis is often much harder than it first appears.
Calibration – The process of building a patient flow model begins with a longitudinal approach, and culminates in a validation of the outputs by comparison with cross-sectional data. This allows the model to be calibrated very accurately to any and all available data, and the greater number of degrees of freedom grants us the flexibility to do it in a credible manner.
Complexity – The process of building a patient flow model begins with a longitudinal approach, and culminates in a validation of the outputs by comparison with cross-sectional data. This allows the model to be calibrated very accurately to any and all available data, and the greater number of degrees of freedom grants us the flexibility to do it in a credible manner.
Data Burden – In order to accurately capture the flow dynamics, you need fairly specific data, all of which is longitudinal in nature:
- Prospective/retrospective/cohort epidemiological studies are available in secondary literature.
- Retrospective data can often be inferred from cross-sectional studies that collect detailed patient history.
- Clinical trial data are widely available.
- Proprietary datasets (insurance claims databases, for example) exist that track patients over time.
If these data don’t exist, it may simply not be possible to quantify all the inputs that the model would require. In our experience, however, this has never proven to be a significant enough problem to derail the effort entirely, although data availability my require you to build your model around the data you have, rather than the data you’d like to have.
Build Time – The process of building a patient flow model begins with a longitudinal approach, and culminates in a validation of the outputs by comparison with cross-sectional data. This allows the model to be calibrated very accurately to any and all available data, and the greater number of degrees of freedom grants us the flexibility to do it in a credible manner.