To illustrate the clinical application of prognostic indices, we highlight 3 common clinical scenarios for older adults in which the consideration of prognosis may improve decision making:
Ms. A is a 78 year old clinic patient with well-controlled diabetes. She has no evidence of cognitive impairment and is independent in all of her activities of daily living. Should you recommend that Ms. A have breast cancer screening?
Ms. B is an 89 year old nursing home patient with congestive heart failure, end stage kidney disease, cognitive decline, and dependence on others for all activities of daily living. Should you discuss the possibility of hospice with her surrogate decision-maker?
Ms. C is an 85 year old with congestive heart failure, chronic kidney disease, and chronic malnutrition who is hospitalized for pneumonia. At the time of discharge she needs help from others to bathe and dress. She has never smoked. An incidental pulmonary nodule was found on admission chest x-ray and confirmed with a CT scan. Should you recommend serial follow-up imaging to her primary care doctor?
Application of prognostic indices proceeds through the steps outlined below.
Step 1: Selecting the Best Prognostic Index (or Indices) for your Patient.
The most appropriate prognostic index for a given patient and clinical situation depends on 5 factors: prediction accuracy, model generalizability, usability, clinical efficacy, and time-frame.
Prediction accuracy refers to how closely the predicted risk of mortality matches the observed risk of mortality for individuals at different risk levels (calibration). Prediction accuracy also considers how well the model separates those at high risk of death from those at low risk of death (discrimination, usually reported quantitatively as a c-statistic). Generalizability refers to how well the index predicts mortality in various populations (e.g., the index was developed from a national sample; does it still work well in my patient population?). Practicality, or usability, refers to models with fewer, easy to enter data elements that are readily available during a clinical encounter. Clinical efficacy refers to testing the index in clinical settings, with a change in clinician behavior and beneficial effects on patient outcomes. None of the indices we examined tested clinical efficacy. Our sorting algorithm provides a potential framework for sorting these factors - see this explanation of how these factors are included in the algorithm.
Finally, clinicians should select indices that predict mortality over a time frame equal to that lag-time to benefit for the intervention. In some cases, no index that matches the time frame is available, but closer scrutiny of existing indices may reveal useful information. We found no index that predicts 6 month mortality for hospitalized elders. However, examination of the risk curve for the Walter 1-year index for hospitalized elders reveals that the highest risk group crosses the 50% predicted mortality mark at 6 months, and may therefore be useful for those seeking to assess hospice eligibility in hospitalized older adults.
Step 2: Estimating Mortality Risk Using a Prognostic Index.
After selecting a prognostic index (or indices) for his patient, the physician can estimate his patient’s mortality by answering a series of questions (i.e. filling in the relevant predictor variables) included in each index. The answers to each question are assigned points and the points are tallied to generate a risk score for each individual. The risk score is correlated with an absolute risk of mortality for a specified time period.
For example, using the Walter index to estimate Ms. C’s risk of dying within 1 year after being discharged from the hospital, one would answer that Ms. C is female (assigned 0 points), needs help with bathing and dressing (2 points), has congestive heart failure (2 points), does not have known cancer (0 points), has an admission creatinine >3.0 (2 points) and albumin <3.0 (2 points). The sum of the points equals a risk score of 8 (0+2+2+0+2+2 = 8), which on the table that Walter provides with her index, correlates with a 64% 1-year mortality for Ms. C. The calculations involved in estimating risk vary in complexity across the selected indices, from simple addition to complex mathematical formulas for others. This online compendium, ePrognosis, uses drop down menus to input predictor values and calculates mortality risk based on published formulas.
Step 3: Interpreting Mortality Risk from a Prognostic Index.
Often clinicians are interested in life expectancy estimates which exceed the time frame of most indices, and life expectancy must be inferred. For example, the Lee index for Ms. A predicts a 6-9% 4-year mortality risk. This means that out of 100 studied subjects with the same risk score as Ms. A, 6 to 9 of them would have died within 4 years. Because median life-expectancy for a group is defined as the point at which 50% of people die, clinicians can infer Ms. A.’s life-expectancy is likely greater than 4 years.
Clinicians must use clinical judgment to decide if the estimated mortality risk is accurate. Clinicians should consider possible sources of error, including limitations of the internal validity of the prognostic index, or error in the application of the index to the specific patient in front of them. Has the index been tested in settings that resemble the patient’s clinical situation? Is the index well calibrated at the predicted risk level – i.e. did observed risk closely match predicted risk in validation studies? Does the index have a good c-statistic, indicating that it accurately stratifies patients into low and high risk groups of mortality? The performance of the index in the original study is likely the ceiling for how well the index will perform in clinical settings. Does the patient have a protective factor (such as social support) that is not captured by the index?
Step 4: Integrating Prognosis into Clinical Care.
We now return to the cases to illustrate the use of prognostic indices in clinical practice:
Using indices developed for community dwelling elders, Ms. A has a 6-9% mortality risk at 4 years using the Lee index, and using the Schonberg index 6% at 5 years and 16% at 9 years. Because of insufficient information in this age group, the US Preventive Services Task Force notes that the balance of benefits and harms of breast cancer screening in women 75 years and older cannot be determined. Life-expectancy and patient preferences therefore play a dominant role in the decision to screen or not screen. Based on a median life expectancy of greater than 9 years, you counsel Ms. A that she is likely to benefit from mammography.
Using an index that predicts 6-month mortality for nursing home patients, Ms. B has a 69% mortality risk at 6 months (Porock index), meaning a life-expectancy of less than 6 months. You counsel her son that Ms. B is eligible for hospice care in the nursing home.
Using an index for hospitalized patients, Ms. C has a 64% one-year mortality risk (Walter index). You suggest to Ms. C's primary care doctor that follow-up imaging for her incidental pulmonary nodule may have limited benefit and even lead to harm from evaluation of false positive findings.