Machine Learning is the study of methods for programming computers to learn. Computers are applied to a wide range of tasks, and for most of these it is relatively easy for programmers to design and implement the necessary software. However, there are many tasks for which this is difficult or impossible. These can be divided into four general categories. First, there are problems for which there exist no human experts. For example, in modern automated manufacturing facilities, there is a need to predict machine failures before they occur by analyzing sensor readings. Because the machines are new, there are no human experts who can be interviewed by a programmer to provide the knowledge necessary to build a computer system. A machine learning system can study recorded data and subsequent machine failures and learn prediction rules. Second, there are problems where human experts exist, but where they are unable to explain their expertise. This is the case in many perceptual tasks, such as speech recognition, hand-writing recognition, and natural language understanding. Virtually all humans exhibit expert-level abilities on these tasks, but none of them can describe the detailed steps that they follow as they perform them. Fortunately, humans can provide machines with examples of the inputs and correct outputs for these tasks, so machine learning algorithms can learn to map the inputs to the outputs. Third, there are problems where phenomena are changing rapidly. In finance, for example, people would like to predict the future behavior of the stock market, of consumer purchases, or of exchange rates. These behaviors change frequently, so that even if a programmer could construct a good predictive computer program, it would need to be rewritten frequently. A learning program can relieve the programmer of this burden by constantly modifying and tuning a set of learned prediction rules. Fourth, there are applications that need to be customized for each computer user separately. Consider, for example, a program to filter unwanted electronic mail messages. Different users will need different filters. It is unreasonable to expect each user to program his or her own rules, and it is infeasible to provide every user with a software engineer to keep the rules up-to-date. A machine learning system can learn which mail messages the user rejects and maintain the filtering rules automatically. Machine learning addresses many of the same research questions as the fields of statistics, data mining, and psychology, but with differences of emphasis. Statistics focuses on understanding the phenomena that have generated the data, often with the goal of testing different hypotheses about those phenomena. Data mining seeks to find patterns in the data that are understandable by people. Psychological studies of human learning aspire to understand the mechanisms underlying the various learning behaviors exhibited by people (concept learning, skill acquisition, strategy change, etc.).

Machine learning

Background
Decision aids are intended to help people participate in decisions that involve weighing the benefits and harms of treatment options often with scientific uncertainty.

Objectives
To assess the effects of decision aids for people facing treatment or screening decisions.

Search methods
For this update, we searched from 2009 to June 2012 in MEDLINE; CENTRAL; EMBASE; PsycINFO; and grey literature. Cumulatively, we have searched each database since its start date including CINAHL (to September 2008).

Selection criteria
We included published randomized controlled trials of decision aids, which are interventions designed to support patients' decision making by making explicit the decision, providing information about treatment or screening options and their associated outcomes, compared to usual care and/or alternative interventions. We excluded studies of participants making hypothetical decisions.

Data collection and analysis
Two review authors independently screened citations for inclusion, extracted data, and assessed risk of bias. The primary outcomes, based on the International Patient Decision Aid Standards (IPDAS), were:

A) 'choice made' attributes;

B) 'decision-making process' attributes.

Secondary outcomes were behavioral, health, and health-system effects. We pooled results using mean differences (MD) and relative risks (RR), applying a random-effects model.

Main results
This update includes 33 new studies for a total of 115 studies involving 34,444 participants. For risk of bias, selective outcome reporting and blinding of participants and personnel were mostly rated as unclear due to inadequate reporting. Based on 7 items, 8 of 115 studies had high risk of bias for 1 or 2 items each.

Of 115 included studies, 88 (76.5%) used at least one of the IPDAS effectiveness criteria: A) 'choice made' attributes criteria: knowledge scores (76 studies); accurate risk perceptions (25 studies); and informed value-based choice (20 studies); and B) 'decision-making process' attributes criteria: feeling informed (34 studies) and feeling clear about values (29 studies).

A) Criteria involving 'choice made' attributes:

Compared to usual care, decision aids increased knowledge (MD 13.34 out of 100; 95% confidence interval (CI) 11.17 to 15.51; n = 42). When more detailed decision aids were compared to simple decision aids, the relative improvement in knowledge was significant (MD 5.52 out of 100; 95% CI 3.90 to 7.15; n = 19). Exposure to a decision aid with expressed probabilities resulted in a higher proportion of people with accurate risk perceptions (RR 1.82; 95% CI 1.52 to 2.16; n = 19). Exposure to a decision aid with explicit values clarification resulted in a higher proportion of patients choosing an option congruent with their values (RR 1.51; 95% CI 1.17 to 1.96; n = 13).

B) Criteria involving 'decision-making process' attributes:

Decision aids compared to usual care interventions resulted in:

a) lower decisional conflict related to feeling uninformed (MD -7.26 of 100; 95% CI -9.73 to -4.78; n = 22) and feeling unclear about personal values (MD -6.09; 95% CI -8.50 to -3.67; n = 18);

b) reduced proportions of people who were passive in decision making (RR 0.66; 95% CI 0.53 to 0.81; n = 14); and

c) reduced proportions of people who remained undecided post-intervention (RR 0.59; 95% CI 0.47 to 0.72; n = 18).

Decision aids appeared to have a positive effect on patient-practitioner communication in all nine studies that measured this outcome. For satisfaction with the decision (n = 20), decision-making process (n = 17), and/or preparation for decision making (n = 3), those exposed to a decision aid were either more satisfied, or there was no difference between the decision aid versus comparison interventions. No studies evaluated decision-making process attributes for helping patients to recognize that a decision needs to be made, or understanding that values affect the choice.

C) Secondary outcomes

Exposure to decision aids compared to usual care reduced the number of people of choosing major elective invasive surgery in favour of more conservative options (RR 0.79; 95% CI 0.68 to 0.93; n = 15). Exposure to decision aids compared to usual care reduced the number of people choosing to have prostate-specific antigen screening (RR 0.87; 95% CI 0.77 to 0.98; n = 9). When detailed compared to simple decision aids were used, fewer people chose menopausal hormone therapy (RR 0.73; 95% CI 0.55 to 0.98; n = 3). For other decisions, the effect on choices was variable.

The effect of decision aids on length of consultation varied from 8 minutes shorter to 23 minutes longer (median 2.55 minutes longer) with 2 studies indicating statistically-significantly longer, 1 study shorter, and 6 studies reporting no difference in consultation length. Groups of patients receiving decision aids do not appear to differ from comparison groups in terms of anxiety (n = 30), general health outcomes (n = 11), and condition-specific health outcomes (n = 11). The effects of decision aids on other outcomes (adherence to the decision, costs/resource use) were inconclusive.

Authors' conclusions
There is high-quality evidence that decision aids compared to usual care improve people's knowledge regarding options, and reduce their decisional conflict related to feeling uninformed and unclear about their personal values. There is moderate-quality evidence that decision aids compared to usual care stimulate people to take a more active role in decision making, and improve accurate risk perceptions when probabilities are included in decision aids, compared to not being included. There is low-quality evidence that decision aids improve congruence between the chosen option and the patient's values.

New for this updated review is further evidence indicating more informed, values-based choices, and improved patient-practitioner communication. There is a variable effect of decision aids on length of consultation. Consistent with findings from the previous review, decision aids have a variable effect on choices. They reduce the number of people choosing discretionary surgery and have no apparent adverse effects on health outcomes or satisfaction. The effects on adherence with the chosen option, cost-effectiveness, use with lower literacy populations, and level of detail needed in decision aids need further evaluation. Little is known about the degree of detail that decision aids need in order to have a positive effect on attributes of the choice made, or the decision-making process.

/pdf/decision-aids-for-people-facing-health-treatment-or-1j1706q17l.pdf

Decision aids for people facing health treatment or screening decisions

OBJECTIVE: Selective serotonin reuptake inhibitors (SSRIs) are widely used to treat depression, but the rates, timing, and baseline predictors of remission in “real world” patients are not established. The authors’ primary objectives in this study were to evaluate the effectiveness of citalopram, an SSRI, using measurement-based care in actual practice, and to identify predictors of symptom remission in outpatients with major depressive disorder. METHOD: This clinical study included outpatients with major depressive disorder who were treated in 23 psychiatric and 18 primary care “real world” settings. The patients received flexible doses of citalopram prescribed by clinicians for up to 14 weeks. The clinicians were assisted by a clinical research coordinator in the application of measurement-based care, which included the routine measurement of symptoms and side effects at each treatment visit and the use of a treatment manual that described when and how to modify medication doses based on these measures....

/pdf/evaluation-of-outcomes-with-citalopram-for-depression-using-53pglcs9e0.pdf

Evaluation of outcomes with citalopram for depression using measurement-based care in STAR*D: implications for clinical practice

Markov models are useful when a decision problem involves risk that is continuous over time, when the timing of events is important, and when important events may happen more than once. Representing such clinical settings with conventional decision trees is difficult and may require unrealistic simplifying assumptions. Markov models assume that a patient is always in one of a finite number of discrete health states, called Markov states. All events are represented as transitions from one state to another. A Markov model may be evaluated by matrix algebra, as a cohort simulation, or as a Monte Carlo simulation. A newer representation of Markov models, the Markov-cycle tree, uses a tree representation of clinical events and may be evaluated either as a cohort simulation or as a Monte Carlo simulation. The ability of the Markov model to represent repetitive events and the time dependence of both probabilities and utilities allows for more accurate representation of clinical settings that involve these issues.

Markov models in medical decision making: a practical guide.

https://www.kidney-international.org/article/S0085-2538(15)48863-5/pdf

Development and progression of nephropathy in type 2 diabetes: the United Kingdom Prospective Diabetes Study (UKPDS 64).

The physician's estimate of prognosis under alternative treatment plans is a principal factor in therapeutic decision making. Current methods of reporting prognosis, which include five-year survivals, survival curves, and quality-adjusted life expectancy, are crude estimates of natural history. In this paper we describe a general-purpose model of medical prognosis based on the Markov process and show how this simple mathematical tool may be used to generate detailed and accurate assessments of life expectancy and health status.

The Markov Process in Medical Prognosis

Background Comorbidity adjustment is an important component of health services research and clinical prognosis. When adjusting for comorbidities in statistical models, researchers can include comorbidities individually or through the use of summary measures such as the Charlson Comorbidity Index or Elixhauser score. We examined the conditions under which individual versus summary measures are most appropriate. Methods We provide an analytic proof of the utility of comorbidity summary measures when used in place of individual comorbidities. We compared the use of the Charlson and Elixhauser scores versus individual comorbidities in prognostic models using a SEER-Medicare data example. We examined the ability of summary comorbidity measures to adjust for confounding using simulations. Results We devised a mathematical proof that found that the comorbidity summary measures are appropriate prognostic or adjustment mechanisms in survival analyses. Once one knows the comorbidity score, no other information about the comorbidity variables used to create the score is generally needed. Our data example and simulations largely confirmed this finding. Conclusions Summary comorbidity measures, such as the Charlson Comorbidity Index and Elixhauser scores, are commonly used for clinical prognosis and comorbidity adjustment. We have provided a theoretical justification that validates the use of such scores under many conditions. Our simulations generally confirm the utility of the summary comorbidity measures as substitutes for use of the individual comorbidity variables in health services research. One caveat is that a summary measure may only be as good as the variables used to create it.

Why Summary Comorbidity Measures Such As the Charlson Comorbidity Index and Elixhauser Score Work.

A convenient approximation of life expectancy (the “DEALE”): II. Use in medical decision-making

Importance Time to surgery (TTS) is of concern to patients and clinicians, but controversy surrounds its effect on breast cancer survival. There remains little national data evaluating the association. Objective To investigate the relationship between the time from diagnosis to breast cancer surgery and survival, using separate analyses of 2 of the largest cancer databases in the United States. Design, Setting, and Participants Two independent population-based studies were conducted of prospectively collected national data from the Surveillance, Epidemiology, and End Results (SEER)-Medicare–linked database and the National Cancer Database (NCDB). The SEER-Medicare cohort included Medicare patients older than 65 years, and the NCDB cohort included patients cared for at Commission on Cancer–accredited facilities throughout the United States. Each analysis assessed overall survival as a function of time between diagnosis and surgery by evaluating 5 intervals (≤30, 31-60, 61-90, 91-120, and 121-180 days) and disease-specific survival at 60-day intervals. All patients were diagnosed with noninflammatory, nonmetastatic, invasive breast cancer and underwent surgery as initial treatment. Main Outcomes and Measures Overall and disease-specific survival as a function of time between diagnosis and surgery, after adjusting for patient, demographic, and tumor-related factors. Results The SEER-Medicare cohort had 94 544 patients 66 years or older diagnosed between 1992 and 2009. With each interval of delay increase, overall survival was lower overall (hazard ratio [HR], 1.09; 95% CI, 1.06-1.13; P P P  = .01). Breast cancer–specific mortality increased with each 60-day interval (subdistribution hazard ratio [sHR], 1.26; 95% CI, 1.02-1.54; P  = .03). The NCDB study evaluated 115 790 patients 18 years or older diagnosed between 2003 and 2005. The overall mortality HR was 1.10 (95% CI, 1.07-1.13; P P P Conclusions and Relevance Greater TTS is associated with lower overall and disease-specific survival, and a shortened delay is associated with benefits comparable to some standard therapies. Although time is required for preoperative evaluation and consideration of options such as reconstruction, efforts to reduce TTS should be pursued when possible to enhance survival.

/pdf/time-to-surgery-and-breast-cancer-survival-in-the-united-4j6tjmphi9.pdf

J. Robert Beck

Papers

Markov models in medical decision making: a practical guide.

The Markov Process in Medical Prognosis

Why Summary Comorbidity Measures Such As the Charlson Comorbidity Index and Elixhauser Score Work.

A convenient approximation of life expectancy (the “DEALE”): II. Use in medical decision-making

Time to Surgery and Breast Cancer Survival in the United States.