SCIENCEYE BLOG

Contact

Machine Learning

26 min read

Updated on November 15, 2026
Painted illustration of glowing data rising from an open ledger into a network of linked lights with a path running between a signpost and a lever

SHARE:

Introduction

Definition

Machine learning is the branch of computing that builds programs which improve automatically through experience.1 Instead of following rules a programmer writes, a machine learning program finds patterns in example data and uses them to make predictions or decisions about new cases. The field lies where computer science meets statistics, and it sits at the core of artificial intelligence and data science.1 Its most powerful modern form is deep learning. A deep model stacks many layers of simple processing units, so it can learn its own representation of raw data such as images or speech: the internal features it uses to describe what it sees or hears.2

Context

The ideas are older than computers. Fitting a line by least squares, regression and the reduction of many measures to a few underlying factors, hidden traits that explain why the measures move together, were all worked out by statisticians long before the first electronic computer. The name arrived in 1959, when Arthur Samuel described a checkers program that improved with practice and put the phrase machine learning in the title of his paper.3 Progress since then has come from new learning algorithms and theory, an explosion of online data and cheap computing power.1

Scope

This entry traces machine learning from its statistical roots to deep learning, sets out its main types and explains how models are built and judged. It then reviews uses in health and exercise science, the difference between prediction and causation, and the field’s main controversies. Causal machine learning, which estimates what would happen under an intervention, is touched on in the section on prediction versus causation.

How a machine learning model learns Four boxes in a loop. Example data with known answers go into a learning algorithm, which adjusts a model. The model makes predictions. Each prediction is compared with the known answer, and the error goes back to the algorithm, which adjusts the model again. A separate arrow shows the finished model making a prediction for a new case. The boxes light up in turn to show the cycle. Learning from examplesExample datawith answersLearningalgorithmModelPredictionCompare withknown answerError adjusts the modelOnce trainedNew caseFinished modelPrediction How a machine learning model learns Four boxes in a loop. Example data with known answers go into a learning algorithm, which adjusts a model. The model makes predictions. Each prediction is compared with the known answer, and the error goes back to the algorithm, which adjusts the model again. A separate arrow shows the finished model making a prediction for a new case. The boxes light up in turn to show the cycle. Learning from examplesExample datawith answersLearning algorithmModelPredictionCompare with theknown answerOrange arrow: the erroradjusts the modelThe finished model thenpredicts new cases
Figure 1. Supervised learning in outline. An algorithm adjusts a model until its predictions match known answers as closely as possible, then the finished model is applied to new cases.1

History

Statistical roots

Much of machine learning still rests on methods statisticians built to summarize data. The method of least squares chooses the line that makes the sum of squared errors between data and model as small as possible. Legendre and Gauss disputed who invented it in the early 1800s.4 In 1886 Francis Galton showed that the children of unusually tall or short parents were, on average, closer to the population mean than their parents, a pattern he called regression.5 The word survives in linear and logistic regression, which remain standard learning algorithms.

In 1901 Karl Pearson showed how to find the lines and planes of closest fit to a cloud of points.6 His idea is the basis of principal component analysis, which compresses many correlated measures into a few. Three years later Charles Spearman argued that scores on different mental tests share one general factor,7 the start of factor analysis. Both are early forms of what is now called unsupervised learning: finding structure in data when there is no outcome to predict.

The perceptron and the name

The first learning machine modeled on the brain appeared in 1958, when Frank Rosenblatt proposed the perceptron. It is a network of simple units whose connection weights change with experience, so it can learn to sort patterns into classes.8 A year later Samuel published his checkers studies. His program learned from its own games, and his title used the phrase machine learning, the name the field has carried since.3

Neural networks fall and return

Rosenblatt’s perceptron had a single layer of adjustable weights. A network with hidden layers between input and output can represent far more complex patterns. Those hidden units are hard to train, though, because no one tells them what they should represent. In 1986 Rumelhart, Hinton and Williams showed how to train them with back-propagation. The method passes the error at the output backward through the network to adjust every weight, and the hidden units learn useful features of the task on their own.9

Other families of algorithms grew alongside neural networks. Random forests, introduced by Leo Breiman in 2001, combine many decision trees, each grown on a random selection of the data and features (the input variables). Their error rates compare favorably with earlier methods that combine many models.10 Random forests remain among the methods most often used in clinical prediction research.11 Support-vector machines, described by Cortes and Vapnik in 1995, took a different route. They map the inputs into a space with many more dimensions and then draw the widest possible boundary between two classes.12

Deep learning

Deep learning is the return of neural networks at scale. A deep model stacks many processing layers, each learning a more abstract representation of the data than the layer below, and it uses back-propagation to tune its internal parameters, the weights it learns.2 By 2015 these methods had greatly improved speech recognition, visual object recognition and object detection, and they had spread into drug discovery and genomics.2 Convolutional networks suit images, and recurrent networks suit sequences such as text and speech.2

A milestone came in 2016. AlphaGo combined deep networks, supervised learning from expert games and reinforcement learning from games against itself, and it defeated the European Go champion by 5 games to 0.13 It was the first time a computer program had beaten a human professional at the full-sized game.13

Milestones in the history of machine learning A timeline of nine milestones: least squares in the early 1800s, regression in 1886, lines of closest fit, the basis of principal component analysis, in 1901, factor analysis in 1904, the perceptron in 1958, the name machine learning in 1959, back-propagation in 1986, random forests in 2001 and the deep learning review of 2015. The first four are statistical roots; the rest are computing milestones. Each milestone lights up in turn. From statistics to deep learningStatistical rootsLearning machines1805Leastsquares1886Regression1901Closest-fitlines (PCA)1904Factoranalysis1958Perceptron1959“Machinelearning”1986Back-propagation2001Randomforests2015DeeplearningWhite: statistical methods. Orange: machine learning milestones. Milestones in the history of machine learning A timeline of nine milestones: least squares in the early 1800s, regression in 1886, lines of closest fit, the basis of principal component analysis, in 1901, factor analysis in 1904, the perceptron in 1958, the name machine learning in 1959, back-propagation in 1986, random forests in 2001 and the deep learning review of 2015. The first four are statistical roots; the rest are computing milestones. Each milestone lights up in turn. From statistics todeep learning1805Least squares1886Regression1901Closest-fit lines (PCA)1904Factor analysis1958Perceptron1959“Machine learning”1986Backpropagation2001Random forests2015Deep learningWhite: statistics. Orange: machine learning.
Figure 2. Milestones from the statistical roots of machine learning to deep learning.2,3,4,5,6,7,8,9,10 The least squares date is Legendre’s publication; Gauss claimed earlier use.

Types of Machine Learning

Machine learning methods are usually grouped by what the data give the algorithm to learn from. Table 1 sets out the main types.

Supervised learning

In supervised learning, each training example comes with the answer, called the label or outcome, and the algorithm learns a rule that maps the inputs to that answer. Classification predicts a category, such as whether a retinal photograph shows diabetic eye disease. Regression, in the machine learning sense, predicts a number, such as blood pressure. Logistic regression, classification trees, random forests, artificial neural networks and support-vector machines are the methods most often used this way in clinical prediction research.11 Most health applications are supervised, because they ask a model to predict a diagnosis or an outcome that has been recorded.

Unsupervised learning

Unsupervised learning has no labels. The algorithm looks for structure on its own: groups of similar people, or a few components that summarize many correlated measures, as Pearson’s and Spearman’s methods did.6,7 Grouping patients by their patterns of risk factors is a typical health use. The groups are descriptions of the data, and they still need testing against outcomes before they carry clinical meaning.

Reinforcement learning

In reinforcement learning, a program learns by acting and receiving rewards or penalties. AlphaGo improved this way by playing against itself.13 The approach suits decisions made in sequence, such as adjusting a treatment over time. Learning by trial and error is hard to arrange safely with patients, however, so the method is harder to apply in health care than in games.

Semi-supervised and self-supervised learning

Between these types sit semi-supervised and self-supervised learning. Both use large amounts of unlabeled data to learn a representation first, and then a smaller labeled set teaches the final task. Deep learning made these approaches practical, because its layers can learn representations directly from raw data.2

Common algorithms

Within each type, a few families of algorithms do most of the work. Linear models, such as linear and logistic regression, give each input a weight and add the results. They are simple to fit and to read, and they are the usual benchmark for other methods.11 Decision trees split the data step by step on the input that best separates the outcomes. Random forests combine the votes of many such trees.10 Support-vector machines find the boundary that leaves the widest margin between classes.12

Neural networks pass the inputs through layers of weighted units. With enough layers and data, they can learn features that no one specified in advance.2,9 No family wins everywhere. The best choice depends on the size of the dataset, the kind of input and how much the user needs to understand how the model reaches its answers.

TypeWhat it learns fromTypical taskHealth example
SupervisedExamples paired with known answersClassify or predict a valueDetecting diabetic eye disease in retinal photographs14
UnsupervisedExamples with no answersFind groups or summary componentsGrouping patients by risk-factor profile
ReinforcementRewards and penalties for actionsChoose a sequence of actionsSimulated treatment strategies
Semi- and self-supervisedMuch unlabeled data, a little labeled dataLearn a representation, then a taskPretraining on unlabeled scans before a labeled task

Table 1. The main types of machine learning.

How Models Are Built and Judged

Building a model

Building a prediction model follows a common sequence, shown in Figure 3. The data are first split: one part develops the model, and another part is set aside, untouched, to test it at the end. Within the development data, the algorithm learns its parameters. The settings that control the learning itself, called hyperparameters, are tuned by cross-validation, which repeatedly trains on part of the development data and checks performance on the rest. In fivefold cross-validation, for example, the development data are cut into five parts. The model is trained five times, each time holding out a different part, and the five results are averaged.

Before any algorithm runs, the team decides which inputs, or features, the model may use. A feature must be available at the moment the prediction will be made. A test result recorded after the outcome makes a model look accurate in development and useless in practice.15 Features also need to be measured the same way where the model will be used. A model can pick up features of the place its data came from, such as a hospital’s equipment, and lose accuracy elsewhere.16

Overfitting is the central danger. A flexible model can fit the noise in its development data and then predict new cases poorly. Riley and colleagues note that many clinical prediction models are developed on datasets too small for the number of participants or outcome events. The result is inaccurate predictions, and they give formulas for the minimum sample size a new model needs.17 The opposite failure, underfitting, happens when a model is too simple to capture a real pattern. Regularization steers between the two. It adds a penalty for complexity, so the model keeps only the patterns strong enough to be worth the cost. The activity classifier described below used this kind of penalized regression.18

A subtler danger is data leakage, in which information about the test cases slips into training. Leakage can happen when the same patient appears in both sets or when a predictor is recorded after the outcome it is meant to predict. Kapoor and Narayanan found leakage in 17 scientific fields, affecting 294 papers and, in some cases, producing conclusions that greatly overstated what the models could do.15

How a prediction model is built and judged A flow of six steps. All data are split into development data and held-out test data. The model is trained and tuned on the development data with cross-validation. It is then tested once on the held-out data, which is internal validation, and later on data from another place or time, which is external validation. Discrimination, calibration and clinical usefulness are reported at each validation step. A warning notes that the test data must never leak into training. Each step lights up in turn. Build, then judge on data the model has not seenAll dataDevelopment data(train and tune)Held-out test data(internal validation)Cross-validationtunes the settingsNew site or period(external validation)Report discrimination, calibrationand clinical usefulnessWarning: test data must never leak into training or tuning. How a prediction model is built and judged A flow of six steps. All data are split into development data and held-out test data. The model is trained and tuned on the development data with cross-validation. It is then tested once on the held-out data, which is internal validation, and later on data from another place or time, which is external validation. Discrimination, calibration and clinical usefulness are reported at each validation step. A warning notes that the test data must never leak into training. Each step lights up in turn. Build, then judge ondata the model has not seenAll dataDevelopment data:train and tune,with cross-validationHeld-out test data(internal validation)New site or period(external validation)Report discrimination,calibration, usefulnessWarning: test data must neverleak into training or tuning.
Figure 3. The usual sequence for building and judging a prediction model. Performance is reported on data the model did not learn from, first held out from the same source and then from a new site or period.15,17,19

Judging a model

A prediction model is judged on three properties, summarized in Table 2. Discrimination is how well it separates people who have the outcome from those who do not. It is usually reported as the area under the receiver operating characteristic curve (AUC), also called the C statistic, where 0.5 is no better than chance and 1.0 is perfect. Calibration is whether predicted risks match observed risks: of 100 people given a 20% risk, about 20 should have the event.

Van Calster and colleagues call calibration the Achilles heel of predictive analytics, because a poorly calibrated algorithm can mislead clinical decisions and is rarely assessed.20 A calibration plot checks calibration by setting predicted risk against observed risk. Clinical usefulness asks whether acting on the predictions leads to better decisions than current practice. Net benefit, often shown as a decision curve, weighs the true cases a model finds against the false alarms it raises. Each property must hold beyond the development data. Internal validation estimates performance on new people from the same source. External validation tests the model in a different hospital, region or period, where patients and measurements differ. The TRIPOD+AI statement sets out what a report of a prediction model should contain, whether the model was built with regression or machine learning.19 Reporting matters because readers cannot judge what they cannot see. In the 71 studies Christodoulou and colleagues reviewed, 79% did not address calibration, and 68% showed potential bias in how the models were validated.11

PropertyQuestion it answersCommon measureWarning sign
DiscriminationDoes the model rank people with the outcome above those without?AUC (C statistic)High AUC only in the development data
CalibrationDo predicted risks match observed risks?Calibration plot and slopePredictions too extreme or too modest20
Clinical usefulnessDoes acting on the model improve decisions?Net benefit, decision curvesMany alerts with little change in care
GeneralizabilityDoes it work in a new place or time?External validationOnly internal validation reported19

Table 2. How a prediction model is judged.

Uses in Health and Exercise Science

Medicine

Imaging was the first clinical area to change. Gulshan and colleagues trained a deep convolutional network on 128,175 retinal images graded by ophthalmologists. In two validation sets it detected referable diabetic retinopathy (moderate or worse disease, needing referral to a specialist) with an AUC of 0.991 (95% confidence interval [CI] 0.988 to 0.993) and 0.990 (95% CI 0.986 to 0.995).14 The authors noted that further research was needed to show whether using the algorithm improves care and outcomes.14

Risk prediction is the second major use. One study followed 378,256 patients from UK general practices. Four machine learning algorithms predicted a first cardiovascular event over 10 years better than an established American College of Cardiology algorithm, whose AUC was 0.728 (95% CI 0.723 to 0.735).21 A neural network reached 0.764 (95% CI 0.759 to 0.769), but logistic regression in the same study reached 0.760 (95% CI 0.755 to 0.766), almost as high.21 The same comparison, a flexible model against regression, applies to vascular measures such as pulse wave velocity.

Exercise science and Physical Activity

Wearable sensors produce the high-volume, repetitive signals that machine learning handles well. Trost and colleagues trained a regularized logistic regression model to recognize seven types of activity, from lying down to dancing, in accelerometer data from 52 children and adolescents. Overall accuracy was 91.0% with the sensor on the hip and 88.4% with it on the wrist.18 Accuracy varied by activity: the wrist model recognized walking 95.8% of the time but dancing only 69.4% of the time.18

In team sports, a systematic review found 58 studies that applied 11 artificial intelligence techniques across 12 sports to predict injury risk or performance. Artificial neural networks, decision tree classifiers and support-vector machines were the most common methods, and 76% of the 6,456 participants were professional athletes.22 The review called for prospective research to establish how well these predictions perform.22

Exercise datasets bring their own risks. Studies often have few participants but many repeated measurements from each one. If windows of data from the same person fall into both the training and the test sets, the model can learn the person rather than the activity, a form of leakage.15 Small samples also limit how many predictors a model can safely use.17

From prediction to practice

Accuracy alone does not show benefit, as the authors of the retinal study noted.14 Topol describes three levels at which artificial intelligence is starting to act in medicine. For clinicians, it mainly speeds image interpretation. For health systems, it can improve workflow and may reduce errors. For patients, it can help them process their own data.23 At each level, a model supports a decision; it does not make the decision. The same holds in exercise science. Activity classifiers turn raw accelerometer signals into minutes of sitting, standing and walking, and those minutes become the exposures in studies of health.18 When a classifier errs on some activities, as it did for dancing, the errors pass into every analysis that uses its output.18

StudySetting and sampleMain findingDOI
Gulshan et al., 2016128,175 retinal images for development; two validation setsAUC 0.991 and 0.990 for referable diabetic retinopathy1410.1001/jama.2016.17216
Weng et al., 2017378,256 UK primary care patientsNeural network AUC 0.764 vs 0.728 for the established algorithm2110.1371/journal.pone.0174944
Trost et al., 201452 children and adolescents with hip and wrist accelerometersActivity recognized with 91.0% (hip) and 88.4% (wrist) accuracy1810.1088/0967-3334/35/11/2183
Claudino et al., 2019Systematic review of 58 team-sport studiesNeural networks, decision trees and support-vector machines most used2210.1186/s40798-019-0202-3

Table 3. Examples of machine learning in health and exercise science.

Prediction Versus Causation

Prediction and causal inference ask different questions of the same data. Breiman described two cultures in statistics. One assumes a known kind of random process generated the data. The other treats the process as unknown and judges algorithms by how accurately they predict.24 Hernán and colleagues divide data science into three tasks: description, prediction and counterfactual prediction (what would happen under a different action), which includes causal inference, and each needs different data, assumptions and methods.25

A model that predicts heart attacks well shows clinicians who is at risk. It does not show what would happen if one of its predictors were changed. Prediction and causation are often confused. In 180 cohort studies from leading medical journals, 26% mixed up causal and predictive aims. Among the prediction studies, 38% did so, most often by reading a predictor as a cause.26 Among the causal studies, 22% were mixed up, mostly because covariates, the other variables a model adjusts for, were chosen for their power to predict rather than from the causal structure.26 Arnold and colleagues list five ways a model built for prediction differs from one built for causal inference. The differences lie in which covariates are considered, how they are chosen, the form they take, how the model is evaluated and how it is interpreted.27 The SciencEye posts Why Machine Learning Can’t Answer ‘What If?’ and Causal Machine Learning: Signposts, Levers and HDL work through examples. Causal machine learning tries to answer that second question: it estimates the effect of a treatment, and how it differs between people, under stated assumptions.28 A directed acyclic graph, a diagram of which variables cause which, is one way to write those causal assumptions down.

A predictor shows where risk is. Only a cause can be changed to lower it.

Controversies and Limitations

Debates

Two debates dominate. The first is whether machine learning beats simpler statistics. A systematic review of 71 studies found no difference between logistic regression and machine learning in the 145 comparisons at low risk of bias (difference in AUC on the logit scale 0.00, 95% CI −0.18 to 0.18, where 0 means no difference). Machine learning looked better only in comparisons at high risk of bias.11 The cardiovascular study above, in which logistic regression came within 0.004 of the best neural network, points the same way.21 When Kapoor and Narayanan corrected leakage errors in a field where complex models were thought to excel, those models did no better than decades-old logistic regression.15 Supporters of machine learning reply that small gains still matter at scale. The neural network in the cardiovascular study correctly identified 355 more people who went on to develop cardiovascular disease than the established algorithm did.21

The second debate is whether models must be explainable. Rudin argues that for high-stakes decisions, models should be interpretable by design, rather than black boxes explained after the fact.29 Ghassemi and colleagues argue that current explanation methods are unlikely to deliver trust, transparency or protection from bias for decisions about individual patients.30 Wider concerns follow the technology into practice. Topol lists bias, privacy and security, and a lack of transparency as its main current limitations in medicine. Whether it will strengthen or erode the relationship between patient and doctor, he writes, remains to be seen.23

Limitations

Performance in development often fails to survive use in practice. The Epic Sepsis Model, a proprietary (commercially owned) tool used in hundreds of US hospitals, reached an AUC of only 0.63 (95% CI 0.62 to 0.64) in an external validation covering 38,455 hospitalizations. It missed 67% of patients with sepsis while generating alerts for 18% of all patients.31 A pneumonia-detection network learned to recognize which hospital system a chest radiograph came from, a shortcut that can confound its predictions. In three of five comparisons it performed worse on outside data than on data from its own hospitals.16

Bias can enter through the outcome a model is trained to predict. A widely used care-management algorithm predicted health-care costs rather than illness. Less money is spent on Black patients with the same level of need, so Black patients were sicker than White patients at the same risk score. Correcting the disparity would raise the share of Black patients receiving extra help from 17.7% to 46.5%.32 Early in the COVID-19 pandemic, all 31 prediction models in a rapid systematic review were rated at high risk of bias.33 The gaps these cases share are external validation, calibration, fair outcome labels and prospective evidence that a model improves care, as Table 4 shows.

The evidence these failures leave behind is uneven. Many models are tested for accuracy, but far fewer are tested for whether using them changes outcomes, the step the retinal and team-sport authors both said was still needed.14,22 Models can also drift. A model calibrated in one period can lose calibration as practice and populations change, and Van Calster and colleagues advise that algorithms be updated when needed to support clinical practice.20 Open questions for the field include how to test models prospectively at reasonable cost, and how to keep them accurate and fair as the data they learned from age. Machine learning is best read as a powerful way to predict from data, and its value in health depends on validation, calibration, fairness and evidence that acting on its predictions helps.

CaseWhat went wrongLesson
Epic Sepsis ModelAUC 0.63 at external validation; missed 67% of sepsis31Validate externally before wide use
Pneumonia radiograph networkLearned the hospital, not only the disease16Check for shortcuts across sites
Care-management algorithmPredicted cost as a stand-in for need32Choose a fair outcome to predict
Early COVID-19 modelsAll 31 at high risk of bias33Follow reporting and appraisal standards

Table 4. Lessons from well-studied failures.

References

  1. Jordan MI, Mitchell TM. Machine learning: trends, perspectives, and prospects. Science. 2015;349(6245):255-60. doi:10.1126/science.aaa8415
  2. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521(7553):436-44. doi:10.1038/nature14539
  3. Samuel AL. Some studies in machine learning using the game of checkers. IBM J Res Dev. 1959;3(3):210-29. doi:10.1147/rd.33.0210
  4. Stigler SM. Gauss and the invention of least squares. Ann Stat. 1981;9(3):465-74. doi:10.1214/aos/1176345451
  5. Galton F. Regression towards mediocrity in hereditary stature. J Anthropol Inst G B Irel. 1886;15:246-63. doi:10.2307/2841583
  6. Pearson K. On lines and planes of closest fit to systems of points in space. Philos Mag. 1901;2(11):559-72. doi:10.1080/14786440109462720
  7. Spearman C. “General intelligence,” objectively determined and measured. Am J Psychol. 1904;15(2):201-92. doi:10.2307/1412107
  8. Rosenblatt F. The perceptron: a probabilistic model for information storage and organization in the brain. Psychol Rev. 1958;65(6):386-408. doi:10.1037/h0042519
  9. Rumelhart DE, Hinton GE, Williams RJ. Learning representations by back-propagating errors. Nature. 1986;323(6088):533-6. doi:10.1038/323533a0
  10. Breiman L. Random forests. Mach Learn. 2001;45(1):5-32. doi:10.1023/A:1010933404324
  11. Christodoulou E, Ma J, Collins GS, Steyerberg EW, Verbakel JY, Van Calster B. A systematic review shows no performance benefit of machine learning over logistic regression for clinical prediction models. J Clin Epidemiol. 2019;110:12-22. doi:10.1016/j.jclinepi.2019.02.004
  12. Cortes C, Vapnik V. Support-vector networks. Mach Learn. 1995;20(3):273-97. doi:10.1007/BF00994018
  13. Silver D, Huang A, Maddison CJ, Guez A, Sifre L, van den Driessche G, et al. Mastering the game of Go with deep neural networks and tree search. Nature. 2016;529(7587):484-9. doi:10.1038/nature16961
  14. Gulshan V, Peng L, Coram M, Stumpe MC, Wu D, Narayanaswamy A, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA. 2016;316(22):2402-10. doi:10.1001/jama.2016.17216
  15. Kapoor S, Narayanan A. Leakage and the reproducibility crisis in machine-learning-based science. Patterns (N Y). 2023;4(9):100804. doi:10.1016/j.patter.2023.100804
  16. Zech JR, Badgeley MA, Liu M, Costa AB, Titano JJ, Oermann EK. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross-sectional study. PLoS Med. 2018;15(11):e1002683. doi:10.1371/journal.pmed.1002683
  17. Riley RD, Ensor J, Snell KIE, Harrell FE Jr, Martin GP, Reitsma JB, et al. Calculating the sample size required for developing a clinical prediction model. BMJ. 2020;368:m441. doi:10.1136/bmj.m441
  18. Trost SG, Zheng Y, Wong WK. Machine learning for activity recognition: hip versus wrist data. Physiol Meas. 2014;35(11):2183-9. doi:10.1088/0967-3334/35/11/2183
  19. Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Van Calster B, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. doi:10.1136/bmj-2023-078378
  20. Van Calster B, McLernon DJ, van Smeden M, Wynants L, Steyerberg EW. Calibration: the Achilles heel of predictive analytics. BMC Med. 2019;17(1):230. doi:10.1186/s12916-019-1466-7
  21. Weng SF, Reps J, Kai J, Garibaldi JM, Qureshi N. Can machine-learning improve cardiovascular risk prediction using routine clinical data? PLoS One. 2017;12(4):e0174944. doi:10.1371/journal.pone.0174944
  22. Claudino JG, Capanema DO, de Souza TV, Serrão JC, Machado Pereira AC, Nassis GP. Current approaches to the use of artificial intelligence for injury risk assessment and performance prediction in team sports: a systematic review. Sports Med Open. 2019;5(1):28. doi:10.1186/s40798-019-0202-3
  23. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44-56. doi:10.1038/s41591-018-0300-7
  24. Breiman L. Statistical modeling: the two cultures. Stat Sci. 2001;16(3):199-231. doi:10.1214/ss/1009213726
  25. Hernán MA, Hsu J, Healy B. A second chance to get causal inference right: a classification of data science tasks. Chance. 2019;32(1):42-9. doi:10.1080/09332480.2019.1579578
  26. Ramspek CL, Steyerberg EW, Riley RD, Rosendaal FR, Dekkers OM, Dekker FW, et al. Prediction or causality? A scoping review of their conflation within current observational research. Eur J Epidemiol. 2021;36(9):889-98. doi:10.1007/s10654-021-00794-w
  27. Arnold KF, Davies V, de Kamps M, Tennant PWG, Mbotwa J, Gilthorpe MS. Reflection on modern methods: generalized linear models for prognosis and intervention: theory, practice and implications for machine learning. Int J Epidemiol. 2020;49(6):2074-82. doi:10.1093/ije/dyaa049
  28. Feuerriegel S, Frauen D, Melnychuk V, Schweisthal J, Hess K, Curth A, et al. Causal machine learning for predicting treatment outcomes. Nat Med. 2024;30(4):958-68. doi:10.1038/s41591-024-02902-1
  29. Rudin C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell. 2019;1(5):206-15. doi:10.1038/s42256-019-0048-x
  30. Ghassemi M, Oakden-Rayner L, Beam AL. The false hope of current approaches to explainable artificial intelligence in health care. Lancet Digit Health. 2021;3(11):e745-50. doi:10.1016/S2589-7500(21)00208-9
  31. Wong A, Otles E, Donnelly JP, Krumm A, McCullough J, DeTroyer-Cooley O, et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Intern Med. 2021;181(8):1065-70. doi:10.1001/jamainternmed.2021.2626
  32. Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447-53. doi:10.1126/science.aax2342
  33. Wynants L, Van Calster B, Collins GS, Riley RD, Heinze G, Schuit E, et al. Prediction models for diagnosis and prognosis of covid-19: systematic review and critical appraisal. BMJ. 2020;369:m1328. doi:10.1136/bmj.m1328

Related Entries

External Resources

Leave a Reply

Your email address will not be published. Required fields are marked *