Prognostic vs. Predictive: How Prognostic Models Work

Prognostic, in medicine, means relating to the likely course or outcome of a disease. When a doctor describes a test result, a tumor characteristic, or a patient feature as “prognostic,” they mean it helps forecast what will happen to the patient over time, regardless of which treatment is chosen. The concept is ancient, but the tools for generating prognostic information have changed dramatically in recent years, from a physician’s bedside impression to complex statistical models, blood-based biomarkers, and machine-learning algorithms. Understanding what prognostic actually means, and where its limits lie, matters for anyone trying to make sense of a diagnosis, a clinical trial report, or a conversation with their care team.

Prognostic Versus Predictive

These two words get swapped constantly, even in medical writing, but they refer to genuinely different things. A prognostic factor tells you something about the natural trajectory of a disease. Stage of cancer, for instance, is prognostic: people diagnosed at earlier stages live longer on average than people diagnosed at later stages, no matter what treatment they receive. A predictive factor, by contrast, tells you whether a particular treatment is likely to work. A gene mutation might be predictive if patients carrying it respond well to a targeted drug while patients without it do not. To prove a factor is truly predictive, researchers need data from patients who received the treatment and patients who did not, ideally from a randomized trial, along with a formal statistical test showing the treatment’s benefit differs depending on the factor’s status.

Some factors are both prognostic and predictive at the same time. A biomarker might signal a worse overall prognosis while also flagging patients who benefit most from a specific therapy. This dual role can create confusion in clinical conversations. If your oncologist says a marker is “prognostic,” they are speaking about the likely course of the disease itself. If they say it is “predictive,” they are speaking about whether a certain treatment is expected to help you specifically.

How Prognostic Factors Are Identified

Prognostic research serves several purposes: describing how a condition typically unfolds, identifying which patient characteristics are linked to better or worse outcomes, estimating an individual’s probability of developing complications, and ultimately informing the design of treatments that could change those outcomes.

² In practice, this means researchers collect data on large groups of patients, record dozens of characteristics at baseline, follow everyone for months or years, and then use statistical methods to determine which characteristics independently predicted who did well and who did not.

In lung cancer, for example, a large analysis found that sex, cancer stage, whether the patient had surgery, and whether they received chemotherapy were all independently linked to survival. Patients with early-stage disease who underwent surgery fared considerably better than those with advanced-stage disease who did not.

³ Similarly, in colorectal cancer after surgery, prognostic models that incorporate multiple patient and tumor features have shown strong agreement between predicted and observed survival outcomes.

⁴

The same general approach applies across medicine, well beyond cancer. During the COVID-19 pandemic, researchers developed a prognostic model called PLANS, built from five simple blood and demographic variables: platelet count, lymphocyte count, age, neutrophil count, and sex. The model performed well in the population it was built from and held up when tested in a separate, independent group of patients.

⁵

What Makes a Prognostic Model Trustworthy

A prognostic model is only useful if it actually works, and “works” has two distinct meanings in this context. The first is discrimination: can the model sort patients into those who will have the outcome and those who will not? The second is calibration: when the model says a patient has a 20 percent chance of dying within a year, do roughly 20 out of every 100 such patients actually die?⁶ A model can be excellent at ranking patients from low to high risk (good discrimination) while still consistently overestimating or underestimating the absolute numbers (poor calibration). Both qualities matter, but calibration tends to get less attention, and poorly calibrated models can lead to bad decisions at the bedside.

A related concern is overfitting. When a model is built on a particular dataset, it can latch onto quirks in that specific group of patients that do not generalize to anyone else. Researchers test for this through internal validation, often using a technique called bootstrapping that simulates how the model would perform in new data by resampling the original dataset hundreds of times.

⁷ Other approaches penalize the model during development itself, forcing it to be less aggressive in fitting every wrinkle of the training data.

⁸ The gold standard, though, remains external validation: testing the model on an entirely separate group of patients from a different time, institution, or country. A model that survives external validation is far more credible than one that has only been tested on the data used to build it.

Why Competing Risks Change the Numbers

One subtlety that even some clinicians overlook is the problem of competing risks. Suppose you want to predict the chance that an older patient will develop liver cancer in the next ten years. If that patient also has advanced heart disease, they might die of a heart attack before they ever develop cancer. Standard statistical models can ignore this reality, effectively pretending that non-cancer death does not exist, and as a result they can substantially overstate the risk of the event you actually care about.

In a study of coronary heart disease prediction among women aged 55 to 90, a standard model classified about 18 percent of individuals as high risk, while a model that properly accounted for the competing risk of dying from other causes classified only about 8 percent as high risk.

⁹ That is a large difference with real consequences: overclassifying people as high risk can lead to unnecessary treatments, testing, and anxiety. For younger, healthier populations, competing risks matter less because the chance of dying from something else in the near term is small. In older or sicker patients, ignoring them can be seriously misleading.

Prognostic Biomarkers in the Bloodstream

One of the most active areas of prognostic research involves circulating tumor DNA, or ctDNA, tiny fragments of tumor genetic material that leak into the blood. Because a simple blood draw can capture these fragments, ctDNA-based tests are sometimes called “liquid biopsies.” Their prognostic value has been demonstrated across several cancer types.

In mantle cell lymphoma, higher ctDNA levels at baseline were linked to shorter survival, and mutations in specific genes like TP53 further worsened the outlook. Patients whose ctDNA became undetectable after treatment had markedly better outcomes, while those with persistently positive ctDNA showed signs of molecular relapse.

¹⁰ In small-cell lung cancer, ctDNA levels changed dynamically during treatment and tracked with progression-free survival.

¹¹

In metastatic breast cancer, researchers identified a ctDNA-based genomic signature that tracked a particular genetic event (loss of the retinoblastoma gene) and was independently associated with worse survival. Patients with the highest scores on this signature had roughly five times the risk of death compared to those with the lowest scores, even after accounting for other known risk factors like the amount of tumor DNA in the blood and the number of metastatic sites.

¹² These findings illustrate how prognostic information is becoming more granular and molecular, moving beyond traditional staging systems toward real-time snapshots of a tumor’s biology.

Machine Learning and Multimodal Prognosis

Artificial intelligence has entered prognostic modeling with considerable hype. The premise is straightforward: machine-learning algorithms can process vastly more variables than a human can hold in mind, potentially uncovering patterns in imaging, lab values, genomics, and clinical notes that traditional statistics would miss. A systematic review of studies combining pathology images with molecular data to predict cancer survival found that multimodal models, those using more than one data type, typically outperformed models relying on a single source.

¹³

That said, the enthusiasm should be tempered. A broad review of multimodal machine learning in healthcare found that while these techniques show potential, their effectiveness depends heavily on the specific data and task involved.

¹⁴ Many of these models have been developed and tested on the same public dataset, and how well they will perform on patients in different hospitals, countries, or healthcare systems remains an open question. The gap between a model’s accuracy in a research paper and its usefulness at the bedside is wider than headlines tend to suggest.

Dynamic Predictions That Update Over Time

Traditional prognostic models produce a single estimate, usually at the time of diagnosis or hospital admission, and that estimate does not change. But patients change. Lab values shift, symptoms worsen or improve, treatments are started and stopped. Dynamic prognostication refers to methods that continuously update a patient’s predicted survival as new information becomes available.

Two main statistical approaches exist for this. One, called landmarking, takes snapshots of a patient’s status at predefined time points and recalculates survival probabilities from each snapshot forward. The other, joint modeling, explicitly links how a patient’s measurements evolve over time with their risk of an event, producing more flexible and theoretically more accurate predictions.

¹⁵ Recent work using flexible versions of joint models has shown improved dynamic prediction performance compared to earlier approaches.

¹⁶ For patients and families living with a serious illness, the ability to get an updated rather than a frozen-in-time prognosis is a meaningful practical advance.

How Accurate Are Clinicians at Prognostication

Before there were models, there were doctors making their best guesses, and the research on clinical prognostic accuracy is sobering. A systematic review of survival predictions in palliative care found extreme variability: the gap between what clinicians predicted and what actually happened ranged from an underestimate of 86 days to an overestimate of 93 days. The majority of studies showed doctors overestimated how long patients would survive.

¹⁷

In terminally ill cancer patients specifically, physicians’ predictions were correct to within one week only about a quarter of the time. They were within four weeks roughly 60 percent of the time. The overestimation bias was consistent: doctors predicted a median survival of 42 days when actual survival was 29 days, and they overestimated by at least four weeks in more than a quarter of cases.

¹⁸ This pattern is not a character flaw; it likely reflects the emotional difficulty of delivering grim news combined with genuine uncertainty. But it has real consequences for end-of-life planning, hospice referrals, and how families spend whatever time remains.

Communicating Prognostic Information to Patients

Even the most accurate prognosis is useless if it is communicated badly. Research on family members’ experiences at the end of life found that nearly half wanted more information about possible outcomes, including knowing that the patient was “sick enough to die.” When prognostic uncertainty was communicated poorly, the damage could last years. Family members described information being cloaked in confusing euphemisms, false hope being offered when it was not warranted, and contradictions between what doctors said and the aggressive treatment being delivered. Those who received clear discussions of prognostic uncertainty, by contrast, reported high satisfaction with care and long-term benefit from having been told the truth.

¹⁹

A separate study of traumatic brain injury found that physicians used one of three strategies: leaving no room for uncertainty, being honest about uncertainty, or presenting a range of possibilities. None of these strategies satisfied the majority of surrogate decision-makers; more than a third were frustrated either by too much ambiguity or by the failure to acknowledge that uncertainty existed at all.

²⁰ Research involving older adults with multiple chronic conditions suggests that a personalized approach, built on a trusting relationship and careful attention to timing, works best. Discussions should be honest within negotiated limits, with clinicians signposting available support and planning for a range of possible futures.

²¹

Bias and Fairness in Prognostic Models

Prognostic models are only as fair as the data they are built on. If historical medical records reflect unequal access to care, biased diagnostic practices, or structural racism, a model trained on those records can perpetuate and even amplify those inequities. A critical review identified four broad categories of concern: whether race should be included as a variable, unequal decision rates across groups, unequal error rates across groups, and potential bias in the outcome the model is trying to predict. The review also warned that some popular approaches to “fixing” algorithm fairness can paradoxically worsen outcomes for everyone.

²²

A concrete example comes from colorectal cancer recurrence prediction. When a model was built without race or ethnicity as a variable, it had substantially higher false-negative rates for Hispanic patients compared to non-Hispanic White patients, meaning it was more likely to miss recurrences in minority patients. Adding race and ethnicity as a predictor actually improved fairness across several metrics.

²³ This finding runs counter to the intuition that “race-blind” models are inherently more equitable. The reality is messier: whether including race helps or hurts depends on the specific context, the outcome being predicted, and the downstream decisions being made.

The Economic Case for Better Prognostic Tools

Better prognostication is not just a clinical nicety; it can save money. In atrial fibrillation, using biomarkers to guide risk prediction and treatment decisions reduced average per-patient costs by about 7 percent while increasing quality-adjusted life years by about 12 percent compared to standard care.

²⁴ In early-stage diabetic kidney disease, a prognostic risk tool that guided medication decisions led to fewer kidney disease events, fewer dialysis starts, and net cost savings per patient.

²⁵ In early breast cancer, adding a gene expression signature to clinical guidelines to decide who really needs chemotherapy saved thousands of euros per patient with no meaningful loss in survival.

²⁶

The common thread across these examples is that more precise prognostic information reduces overtreatment. When you can identify who is genuinely at high risk and who is not, you spare low-risk patients from expensive, potentially harmful interventions they did not need. That saves both suffering and dollars.

Reporting Standards and Why They Matter

Much prognostic research has been poorly conducted and interpreted, and the field has responded with formal reporting standards. The TRIPOD statement provides a 22-item checklist aimed at improving how studies developing, validating, or updating prognostic models are reported.

²⁷ Additional tools like the PROBAST risk-of-bias assessment and the CHARMS checklist help reviewers evaluate whether a prognostic study’s methods are sound.

²⁸

For the general reader, the practical takeaway is that not all prognostic tools are created equal. If a doctor mentions a prognostic score or model, reasonable questions include: Has it been validated in patients like me? Was it tested outside the institution where it was built? And does it account for all the relevant outcomes, not just the one the researchers found most interesting? These standards exist precisely because the history of prognostic modeling is littered with tools that looked promising in development and fell apart in practice.

Getting Prognostic Tools Into Everyday Clinical Use

Even well-validated prognostic models often gather dust. Despite the wide availability of risk prediction tools, relatively little has been done to integrate them into electronic health records where they could actually inform real-time decisions.

²⁹ The barriers are partly technical, involving interoperability standards and software development, and partly human. Clinicians may not trust the model, may not understand its output, or may not have time to interpret it during a ten-minute appointment. Regulatory frameworks are also evolving; AI-based prognostic tools classified as medical devices face requirements around quality management, cybersecurity, and lifecycle control that span multiple jurisdictions.

³⁰

How prognostic information is presented to patients and families matters as much as the information itself. Research on patient decision aids has shown that different numerical formats can bias how people perceive risk, potentially nudging them toward one option without meaning to.

³¹ Saying “you have a 90 percent chance of surviving” feels psychologically different from “one in ten people with your condition will die,” even though the numbers are identical. Designing prognostic tools that communicate clearly, not just calculate accurately, remains one of the field’s hardest unsolved problems.