Prognosis is the assessment of the likely future course and outcomes of a health condition. It may concern recovery, recurrence, complications, functional ability, symptoms, or survival. Whereas diagnosis identifies a condition and etiology concerns its causes, prognosis asks what is likely to happen over a specified period, given what is currently known. It is usually expressed as a probability, a range of possible outcomes, or a qualitative judgment rather than as a certain prediction.
Prognosis is central to medicine because the consequences of a condition often matter as much as its identification. Prognostic information informs treatment discussions, clinical research, planning of services, and patients’ understanding of their illness. Its interpretation depends on the outcome being predicted, the starting point and duration of follow-up, the population studied, and the care available. A prognosis is therefore conditional: it describes outcomes under particular circumstances, not an unalterable destiny.
Meaning, scope, and historical development
The word prognosis derives from Greek terms meaning foreknowledge. In the Hippocratic medical tradition, especially the treatise commonly known as Prognostic, anticipating the course of illness was an important part of medical practice. Physicians observed changes in breathing, sleep, bodily appearance, excretions, and other signs to judge whether a patient was likely to recover or deteriorate. Such judgments preceded modern knowledge of disease mechanisms and often concerned the immediate course of an acute illness rather than a precisely defined long-term probability.
The development of vital statistics, life tables, and systematic clinical observation gradually gave prognosis a numerical basis. John Graunt’s analysis of London mortality records in 1662 and Edmond Halley’s life table published in 1693 were important early contributions to population-level analysis of survival. These were not individualized clinical prediction systems, but they helped establish that patterns of death and survival could be studied quantitatively. Later hospital records, disease registries, and prospective studies made it possible to relate clinical characteristics to subsequent outcomes.
Modern prognostic research encompasses several distinct questions. Research on overall prognosis describes what happens to a defined group of people with a condition. Prognostic-factor research examines whether a particular characteristic is associated with a later outcome. Prognostic-model research combines several characteristics to estimate risk for an individual. Research on variation in treatment effects asks a different question: whether the benefits or harms of an intervention differ between people with different characteristics.
This distinction matters because predicting an outcome is not the same as explaining it or identifying the best intervention. A characteristic may be useful for forecasting without being a cause of the outcome. For example, a laboratory measurement may reflect existing organ damage rather than contribute to it. Altering the measurement itself would not necessarily improve health. Conversely, a causal factor may add little predictive information if its effects are already captured by other variables.
The everyday terms good, poor, favorable, and guarded prognosis are less precise than a numerical estimate. “Guarded” generally conveys substantial uncertainty or concern about serious outcomes, but it has no universal numerical definition. A prognosis becomes more interpretable when it specifies the event, time horizon, clinical context, and degree of uncertainty. The same person may have a favorable short-term survival prognosis but an unfavorable outlook for complete functional recovery, so a single descriptive label can conceal important differences.
Outcomes, time horizons, and the meaning of recovery
A prognosis requires a defined outcome. Death is relatively clear as an event, but even survival can be measured in several ways. Overall survival concerns death from any cause. Cause-specific survival concerns death attributed to a particular disease. Recurrence-free, disease-free, and progression-free survival use different combinations of events, and their definitions can vary between studies. In cancer research, for example, progression-free survival generally concerns the time until disease progression or death, whereas overall survival concerns death regardless of whether progression was previously recorded.
Survival is only one dimension of outcome. After stroke, relevant outcomes include mobility, communication, cognition, independence, and the ability to return to previous activities. In chronic illness, symptom burden, frequency of hospital admission, and health-related quality of life may be central. A patient-reported outcome captures aspects such as fatigue, pain, or perceived functioning directly from the person experiencing them. These outcomes can diverge: a treatment may prolong life without restoring function, or improve symptoms without changing survival.
The time horizon fundamentally changes the question. The probability of deterioration within 24 hours is not interchangeable with the probability of death within one year. Acute-care prediction often uses horizons measured in hours or days, whereas chronic-disease models may estimate events over five or ten years. A model’s starting point also matters. Estimates made at diagnosis, after an operation, or following a period of remission refer to different populations and different clinical states.
“Recovery” requires an operational definition. It may mean eradication of an infection, normalization of a laboratory result, resolution of symptoms, restoration of independence, or return to a previous role. These events need not occur together. Remission means a reduction or disappearance of disease manifestations according to specified criteria; it does not invariably mean cure. Cure may refer clinically to the expectation that disease will not return, or statistically to a group whose excess mortality eventually becomes negligible relative to an appropriate comparison population.
Several commonly quoted numbers answer different questions. A five-year survival proportion describes the estimated fraction alive five years after a defined starting point. Median survival is the time at which estimated survival reaches 50%; it is not a deadline for an individual. Life expectancy is an average remaining lifetime under specified assumptions, not a most likely date of death. A complete description of prognosis may therefore require several outcomes and time horizons rather than one headline statistic.
Biological, functional, and social determinants
Prognosis depends partly on the underlying disease: its type, extent, activity, and biological behavior. In oncology, cancer staging describes the anatomical extent of disease, while pathological grade describes features related to differentiation and aggressiveness. The presence of metastases often changes both expected outcomes and available treatments. Molecular characteristics can further distinguish diseases that appear similar under the microscope. Disease labels therefore define broad groups, within which outcomes may vary substantially.
A person’s physiological reserve also influences outcome. Age, organ function, nutritional status, and coexisting conditions can affect the ability to withstand illness or treatment. In heart failure, prognostic information may include symptoms, kidney function, blood pressure, exercise capacity, and biochemical measurements. In serious infection, the extent of organ dysfunction can be more informative about immediate risk than identification of the organism alone. The significance of a characteristic depends on the disease and the time horizon; there is no universal list of predictors with fixed importance.
Functional status often provides information not captured by a diagnosis. Performance-status scales in cancer describe activity and self-care, while frailty assessments examine vulnerability associated with reduced reserve across multiple systems. Frailty is not synonymous with chronological age or with having several diseases. Similarly, dependence in daily activities may arise from neurological impairment, musculoskeletal disease, cognitive difficulties, or environmental barriers. These distinctions affect how a prognostic association is interpreted.
A biomarker can provide information about a biological process, disease severity, or response to treatment. However, prognostic and predictive biomarkers have different technical meanings. A prognostic biomarker is associated with the likelihood of an outcome in a defined clinical setting. A predictive biomarker identifies differences in the effect of a particular treatment. A marker can serve both functions, but evidence for one does not establish the other. For example, a tumor characteristic associated with recurrence does not automatically establish that a particular drug is especially effective in tumors with that characteristic.
Prognosis is also shaped by social determinants of health, including housing, income, education, working conditions, social support, and access to services. A biological condition may have different observed outcomes where timely diagnosis, medication, rehabilitation, or follow-up are less available. Such differences are not necessarily intrinsic properties of the affected populations. They may reflect modifiable features of the health system and wider society. Prognostic models built from these outcomes can consequently encode both disease processes and the circumstances in which care was delivered.
Sources of prognostic evidence
The central design for studying prognosis is the cohort study, in which a defined group is followed over time. A useful cohort has clear eligibility criteria, a clinically meaningful starting point, consistent measurements, and sufficiently complete follow-up. An inception cohort recruits participants at a common early stage, such as first diagnosis or first presentation. This reduces the distortion that occurs when a study includes only people who have already survived a variable period of illness.
Prospective studies define measurements and outcomes before follow-up occurs. Retrospective studies reconstruct a cohort from existing records. Either can yield valuable information, but they have different vulnerabilities. Routine records may lack symptom severity, functional status, or outcomes occurring outside the recording system. A specialist referral center may receive unusually complex cases, while a community cohort may include milder disease. Consequently, prognosis observed in one setting may not transfer directly to another.
Disease registries and linked administrative data can provide large samples and long follow-up. Population-based cancer registries, for example, help estimate survival across broad geographical areas rather than only among people treated at a particular hospital. An electronic health record can supply repeated laboratory results, diagnoses, and medication histories. However, recorded data reflect clinical practice as well as illness: a test may be ordered more often when deterioration is suspected, so the presence or timing of a measurement can itself contain information about clinicians’ concerns.
A randomized controlled trial provides strong evidence about the effects of the interventions being compared, but its participants may not represent everyone encountered in practice. Eligibility restrictions can exclude people with substantial comorbidity, advanced age, or impaired organ function. Follow-up and treatment adherence may also differ from routine care. Prognostic estimates from a trial therefore describe outcomes under the trial’s treatment and monitoring conditions, not necessarily the untreated course or the experience of the whole population.
Evidence quality depends on more than sample size. Outcome definitions, measurement reliability, missing data, loss to follow-up, and the number of observed events all matter. A large dataset cannot automatically correct systematic bias. Systematic reviews can compare findings across cohorts, but combining studies with different starting points, treatments, or endpoint definitions may produce a misleading average. Prognostic evidence is most interpretable when the target population and clinical setting are explicitly defined, and when uncertainty and differences between studies remain visible.
Survival analysis, censoring, and competing events
Many prognostic questions concern the time until an event. Survival analysis provides methods for these questions, including situations in which not everyone has experienced the event by the end of observation. Let be the time to an event. The survival function is
the probability that the event has not occurred by time . When death from any cause is the event, this is the probability of remaining alive beyond that time.
Follow-up is often incomplete. Someone may still be alive when the study ends, or may leave observation before their eventual outcome is known. Such observations are right-censored: the event time is known to exceed the last observed time, but its exact value is unknown. The Kaplan–Meier estimator, introduced by Edward Kaplan and Paul Meier in 1958, estimates survival by combining the conditional probabilities of surviving successive event times. Its validity depends on assumptions about censoring, including that censoring is not informative about subsequent event risk in ways unaccounted for by the analysis.
The hazard is the instantaneous event rate among those who have not yet experienced the event. It is not the same as the probability of an event over a fixed period. The Cox proportional hazards model, introduced by David Cox in 1972, represents the hazard as
where is a baseline hazard and the variables are predictors. In its standard form, it assumes that hazard ratios remain constant over time. A hazard ratio alone does not provide an individual’s absolute risk; estimating that risk also requires information about baseline survival and the predictor values.
A competing risk is an event that prevents the event of interest from occurring. Death from another cause competes with death from a specified disease. Likewise, death before kidney failure prevents the later occurrence of kidney failure. Treating competing events simply as ordinary censoring can overestimate the probability of the event of interest when Kaplan–Meier methods are used to estimate that probability. Cumulative-incidence methods explicitly account for the fact that competing events remove the possibility of a subsequent first event of the specified type.
Conditional survival updates the question after a person has already survived a period of time. For an additional interval , given survival to time ,
This explains why the outlook several years after diagnosis may differ from the outlook at diagnosis. Recurrent-event and multistate models extend these methods to repeated admissions, relapse, recovery, and transitions between disease states rather than treating every clinical course as a single irreversible event.
Prognostic models and their evaluation
A clinical prediction model combines multiple characteristics to estimate the probability of an outcome or the time until it occurs. Some models use a small set of routinely available measurements and produce a point score; others use regression equations or complex computational methods. A model’s intended population, moment of use, predictors, outcome, and prediction horizon define its purpose. A score developed for intensive-care patients cannot be assumed to work for people assessed in primary care.
Model development involves choosing and measuring predictors, handling missing values, estimating relationships, and evaluating performance. Continuous variables such as age or blood pressure contain more information than crude categories, and their relationships with outcomes may be nonlinear. Arbitrary cutoffs can discard information and create artificial jumps in predicted risk. Overfitting occurs when a model captures accidental features of its development data rather than patterns that generalize. It is especially likely when many parameters are estimated from relatively few outcome events.
Two major aspects of performance are discrimination and calibration. Discrimination is the ability to distinguish people who experience an outcome from those who do not, or to rank earlier events above later ones. Measures include the area under a receiver-operating-characteristic curve and concordance statistics. Calibration concerns agreement between predicted probabilities and observed event frequencies. If a group assigned a 20% risk experiences the event in about 20% of cases, that is evidence of calibration for that group and horizon. A model can discriminate well while systematically overestimating everyone’s risk.
Internal validation assesses optimism within the development sample, commonly through bootstrapping or cross-validation. External validation evaluates performance in new participants, such as people from another hospital, region, or period. Validation is not a permanent certification. Changes in treatment, disease frequency, referral patterns, or measurement methods can alter performance. Recalibration may update a model’s baseline risk or the strength of its predictions, while more substantial changes may require revising the model itself.
Machine learning can identify complex patterns in large datasets, but it does not remove the need for appropriate study design, reliable outcomes, or external validation. Data leakage—using information that would not actually be available at the intended prediction time—can produce deceptively strong results. Measures such as the Brier score assess aspects of probabilistic accuracy, while decision-curve analysis examines potential clinical net benefit across decision thresholds. Even strong statistical performance does not establish that using a model improves health outcomes; that requires evaluation of how it changes decisions and care.
Prognosis across major clinical settings
Cancer prognosis illustrates the interaction of anatomy, biology, treatment, and competing causes of death. Stage describes disease extent, but cancers at the same stage can differ in grade, molecular features, and treatment sensitivity. In breast cancer, hormone-receptor and HER2 status help characterize biologically distinct groups and inform treatment selection. Outcomes also depend on age, comorbidity, treatment received, and access to follow-up. Survival estimates from older cohorts may lag behind outcomes after major therapeutic advances, particularly when long follow-up is required.
In cardiovascular disease, models may estimate either the first occurrence of an event or outcomes after disease is established. Those are different tasks. A model predicting a first myocardial infarction in people without known cardiovascular disease is not equivalent to a model predicting death after an infarction. In chronic heart failure, risk changes after hospitalization, alteration of treatment, or deterioration in kidney function. A single baseline estimate may therefore become less informative as the clinical state changes.
Chronic kidney disease provides another example of multiple relevant outcomes. Prognostic models may estimate progression to kidney failure, but death can occur before kidney failure, particularly in older populations with other illnesses. Estimated filtration rate and urinary albumin measurements provide different information about kidney damage and function. A prediction of kidney failure requiring treatment also depends partly on how that endpoint is defined and on local treatment practices; it is not identical to a purely biological threshold.
In acute illness such as sepsis, prediction often concerns short-term death or deterioration. Physiological abnormalities and the degree of organ dysfunction can change rapidly, so the timing of measurement is crucial. Intensive-care severity scores such as APACHE and SAPS were developed partly to characterize patient groups and compare observed with expected outcomes. A score that performs well for risk adjustment across large groups may still provide an imprecise estimate for one person. It also does not directly establish the benefit or futility of a particular treatment.
Neurological and mental-health conditions often require multidimensional outcomes. After severe brain injury, survival, consciousness, cognition, and independence are distinct questions. In dementia, progression can be described through changes in cognition, daily functioning, complications, and care needs. In depression, relevant outcomes include symptom remission, recurrence, functioning, and quality of life. Across these settings, rehabilitation and environmental support may alter participation and independence even when the underlying lesion or disease persists. Prognostic statements centered exclusively on mortality can therefore omit outcomes of major importance to patients and families.
Updating prognosis and recognizing systematic distortion
Prognosis changes as new information becomes available. The course after initial treatment, the appearance of complications, and repeated measurements can be more informative than baseline characteristics alone. Dynamic prediction uses information accumulated over time to revise estimates. Landmark analysis makes predictions among people who are event-free at a selected time, using information available by that point. Joint models can relate repeated measurements, such as biomarker trajectories, to time-to-event outcomes while accounting for their statistical dependence.
Observed prognosis is also influenced by treatment. People at greater baseline risk may receive more intensive care, reducing the difference in outcomes between high- and low-risk groups. Conversely, frail patients may receive less intensive treatment because of contraindications or treatment preferences. Such patterns complicate the interpretation of associations between treatment, predictors, and outcome. Predicting what happened under existing clinical practice is not the same as estimating what would happen under an alternative intervention; the latter requires methods addressing causal inference.
Screening can change apparent prognosis without postponing death. Lead-time bias occurs when earlier diagnosis lengthens measured survival from diagnosis even though the time of death is unchanged. Length bias arises because screening is more likely to detect slowly progressing disease with a longer detectable preclinical phase. Overdiagnosis adds conditions that would never have caused symptoms or death. These effects help explain why improved survival after diagnosis is not, by itself, proof that a screening program reduces mortality.
Selection can introduce further distortion. Hospital cohorts exclude people who never reach hospital, and studies beginning after treatment exclude those who died or deteriorated before treatment began. Immortal-time bias arises when a period during which a participant must remain alive to qualify for a group is incorrectly attributed to that group. Missing outcomes can also mislead if people lost to follow-up have a different prognosis from those retained. The direction and size of these biases depend on the study design rather than on sample size alone.
Finally, a prognostic estimate may contribute to the outcome it predicts. In some settings of severe neurological injury, predictions can influence decisions about life-sustaining treatment. If the subsequent deaths are then used to validate the prediction without considering those decisions, a self-reinforcing pattern can result. Similar issues arise when high-risk labels alter access to procedures or services. Prognostic research consequently distinguishes outcomes associated with the disease from outcomes partly shaped by responses to the prediction.
Communication, decision-making, and ethical limits
Prognostic communication translates population evidence into information relevant to an individual. Numerical estimates describe frequencies among people with comparable measured characteristics; they do not identify which individual will experience an event. Uncertainty arises both because outcomes vary even among apparently similar people and because knowledge is incomplete. A narrow confidence interval around a population average does not imply a narrowly predictable personal outcome.
The way information is presented affects understanding. “A 90% chance of surviving one year” and “a 10% chance of dying within one year” describe complementary outcomes when the same definitions apply, but they can evoke different responses. Absolute risks, a clear time horizon, and consistent denominators make comparisons more interpretable than relative changes alone. A reduction from 20% to 10%, for example, is a 10-percentage-point absolute reduction and a 50% relative reduction. Neither figure is meaningful without specifying the event and circumstances.
Prognosis contributes to shared decision-making, but it does not determine a choice by itself. Two people with similar predicted outcomes may value longevity, independence, symptom relief, or treatment burden differently. Moreover, the baseline probability of an event and the effect of an intervention are separate quantities. High risk does not automatically imply large treatment benefit, and low risk does not establish the absence of benefit. Treatment effects, harms, uncertainty, and the person’s priorities remain distinct elements of deliberation.
In serious illness, palliative care addresses symptoms, communication, and support alongside other treatment when appropriate. Its role is not restricted to people with a precisely predictable short survival. Prognostic discussions may concern possible functional decline, increasing care needs, or recurrent crises as well as the risk of death. Such communication is distinct from assigning certainty to a life expectancy or treating an estimate as a fixed boundary.
Ethical concerns also arise when predictions influence access to treatment, insurance, or scarce resources. A model can reproduce inequities if historical differences in care are treated as immutable patient characteristics. Evaluation therefore includes performance across relevant groups and the consequences of errors, not merely an overall accuracy measure. When genetic information or large clinical datasets are involved, data privacy and appropriate governance are additional concerns. A prognostic model estimates defined outcomes under specified conditions; it does not measure a person’s worth, the value of their life, or the legitimacy of their preferences.