Course literature

Question

What is data science, and why should scientists care about it? The authors look at data science from three perspectives (statistical, computational and human) and argue that combining all three well is what data science is really about.

Why it matters

Scientists in many fields now have huge amounts of data but cannot yet use it fully. Examples are genetic data linked to people's diseases, large archives of digitised texts for social scientists, and sky surveys in astronomy with hundreds of terabytes of images. The existing methods from statistics and computing are not built for these modern problems: very large datasets, data with very many variables, models of the world that are always somewhat wrong ('misspecified'), and finding out what causes what. The authors see this tension as the trigger for the new label 'data science'.

Approach

This is an opinion essay; it contains no data. It builds on Tukey (1962), who described 'data analysis' as much broader than mathematical statistics. The statistical perspective covers uncertainty (every dataset has some), complex data with links over time, space or between variables (handled for example with Bayesian models: models that state assumptions about the world as probabilities and update them with the data), data with thousands of variables per person (handled with regularisation, which keeps a model simple so it does not chase noise, and with machine learning such as deep learning for prediction), and causality (telling cause apart from correlation). The computational perspective covers how methods run as algorithms on a computer: optimisation (finding the best model settings by stepwise 'climbing' towards the best fit, for example the highest likelihood, i.e. how probable the data are under the model), sampling methods (the bootstrap: drawing many resamples from the data to estimate a confidence interval; MCMC (Markov chain Monte Carlo): a sampling method for Bayesian models) and distributed computing (spreading data and work over many computers). The human perspective is shown with a neuroscientist who films mice brains and behaviour and works with a data scientist.

Findings

• Data science is the 'child of statistics and computer science': it takes over their methods and blends, refocuses and develops them for modern scientific data, in the spirit of Tukey's broad 'data analysis'.
• Statistical perspective: all datasets involve uncertainty, and statistics is the foundation for reasoning about it. Three key areas are complex, structured data (e.g. Bayesian models), high-dimensional data, meaning very many variables per data point (regularisation; machine learning such as deep learning for prediction), and causality (correlation is not causation; drawing conclusions from observational data, where nothing was controlled by the researcher).
• Computational perspective: how methods are run as algorithms, and the trade-off between statistical accuracy and computer resources (time and memory). Examples are optimisation, sampling (the bootstrap for confidence intervals, MCMC for Bayesian models) and distributed computing.
• Human perspective: data science cannot be fully automated, because applying the tools needs human judgement and deep knowledge of the field. The data scientist works step by step together with the domain expert (the expert in the field, e.g. a coach or physician; data scientist and expert can be one person wearing two 'hats'). The work cycles through preprocessing, exploration, selection, transformation, analysis, interpretation and communication. Reproducibility (others can repeat the analysis and get the same result) and data provenance (a record of where the data came from and what was done to them) matter.
• Conclusion: data science is more than statistics plus computer science. It means weaving both into a larger framework, problem by problem; understanding the context of the data; taking responsibility for private and public data; and communicating clearly what a dataset can and cannot tell us.

Question

The authors want to give sport business professionals, coaches, policy makers and other non-technical readers a broad overview of the AI and machine learning methods used for sport performance and sport business problems. AI (artificial intelligence) means computer systems that do tasks that normally need human intelligence; machine learning (ML) is the part of AI where the computer learns patterns from data. The authors note that for many non-experts the link between AI and sport is still 'fuzzy' and the reasons for using ML are unclear. They also discuss how AI could shape sport in the coming years.

Why it matters

Sport is using AI more and more, driven by more computing power and more data. It started with Moneyball and SABRmetrics in baseball (using in-game statistics to find undervalued players) and now includes injury models, player tracking and ticket pricing. Decision-makers need to understand how data scientists think, so they can discuss the approach and method with them without needing the technical details.

Approach

No data are collected; it is a narrative essay. The authors first summarise earlier work on AI in sport: expert systems (programs built from rules written by experts), artificial neural networks and deep learning (models loosely inspired by the brain that learn complex patterns from many examples), evolutionary computation (methods that 'breed' better solutions, mimicking natural selection), Bayesian approaches (probability models updated with data), and the survey by Beal et al. (2019). They list four application areas. They then contrast classic analytics (rules + data → program → answers) with the ML way of working (data + known answers → the algorithm works out the rules itself, which are then tested on new data it has never seen). They write an ML prediction as f(w1·i1, …, wn·in) = y: each input i gets a weight w, and the model combines them into an output y. Next come three made-up sport examples: supervised injury prediction in Australian football, unsupervised fan grouping with K-means clustering (an algorithm that puts similar data points in the same group, where k is the number of groups), and reinforcement learning (Q-learning, SARSA: algorithms that learn by trial and error from rewards) for fantasy sport and game-playing AI, plus a short note on genetic/evolutionary algorithms. They end with views on the future of AI in sport.

Findings

• Machine learning turns classic analytics around: instead of programming the rules, you give the algorithm data together with the known answers, it learns the rules, and those rules are checked by testing accuracy on new (unseen) data.
• Supervised learning (the outcome is known in the past data): e.g. for each Australian football player you know whether he got injured and missed the next match, plus his match load, metres run, warm-up and tackles. The model learns from this, is tested on unseen data and is adjusted until accuracy is good enough (their example figure: 70%). Neural networks, decision trees or regression models can all do this.
• Unsupervised learning (no outcome labels): the algorithm finds patterns nobody had noticed. Example: a football club groups its stadium visitors by gender, age, postcode or income with K-means. The groups come from the data, although the number of groups can be chosen in advance.
• Reinforcement learning: an agent (the decision-making program) acts in a (simulated) environment, gets rewards or penalties, and learns a strategy that collects the most reward over time. Examples are picking fantasy sport teams and game AI (chess, Go, poker, StarCraft).
• AI/ML in sport covers four areas: game activity and analytics, talent identification and recruitment, training and coaching, and fan- and business-focused applications.
• Future: AI is unlikely to fully replace coaches and human experts. The main barriers to complete '360-degree' analyses of an athlete's value are that player, team and commercial data are kept private by their owners (proprietary) and privacy rules. Who owns the data, and ethics (e.g. a team's injury-prediction model versus a player's right to his own data), will become contested.

Question

Which body and muscle characteristics, measured from the whole body down to single muscle fibres, explain sprint performance, endurance performance and the ability to combine both? The underlying problem: large muscle fibres give power for sprinting, and fibres with a high oxidative capacity (the ability to use oxygen to make energy) give endurance, but large fibres tend to have a low oxidative capacity. So improving both at once is hard.

Why it matters

Many sports need both sprint and endurance. Earlier studies looked at only one of the two, and mostly at whole-body measures. Training for endurance and training for peak power interfere with each other, partly inside the muscle cells. Knowing the critical determinants (the factors that matter most) gives targets for training. The two performance tests plus a physiological profile could also help with talent identification and individual training plans.

Approach

This was a cross-sectional study: each cyclist was measured once, with no follow-up over time. It included 28 cyclists (14 road, 8 team pursuit, 6 track sprint), all at national to Olympic level except 4 amateur road cyclists. Sprint performance was the highest 1-second power in a Wingate test (a 30-second all-out sprint on a bike). Endurance performance was the average power in a 15-km time trial. Both were divided by lean body mass (body weight without fat) to the power 2/3, to remove the advantage that bigger riders have just from being bigger. The authors measured many possible determinants: whole-body oxygen uptake (VO2) during the time trial ('performance VO2'), maximal oxygen uptake (VO2max) and thresholds; gross efficiency (the share of the body's energy use that ends up as power on the pedals); blood values such as haemoglobin (Hb, the oxygen-carrying protein in red blood cells) and MCHC (how much haemoglobin is packed into the red blood cells); muscle oxygenation (how much oxygen the thigh muscle holds, measured with a light sensor on the skin, near-infrared spectroscopy); maximal knee-extension force; 3D ultrasound of the outer thigh muscle (vastus lateralis) for muscle volume, fascicle length (length of the fibre bundles), PCSA (physiological cross-sectional area: how thick the muscle is measured across its fibres) and fibre angle; and a muscle biopsy (a small piece of muscle taken with a needle), stained to measure fibre type (fast or slow), fibre size, oxidative capacity (with the enzyme SDH), capillaries (tiny blood vessels that bring oxygen) and myoglobin (the protein that stores oxygen in muscle). To score how well each rider combined sprint and endurance, they fitted a line through all riders' sprint and endurance scores with Deming regression (a straight-line fit that allows for measurement error in both variables, not only in the y-variable). Each rider's score was his perpendicular distance to that line (his residual): above the line means better at combining both than the group trade-off predicts. Links were tested with Pearson correlations (r: the strength of a straight-line relationship, from −1 to 1) and stepwise multiple regression: predictors were added one at a time and kept only if they significantly raised R² (the share of the differences between riders that the model explains, from 0 to 1; P < 0.05).

Findings

• Sprint and endurance performance (corrected for body size) were inversely related: r = −0.66 (P < 0.001); per kg body mass r = −0.44. So being good at both is difficult.
• Sprint: the percentage of fast fibres and the volume of the thigh muscle together explained 65% of the differences in sprint power (R² = 0.65).
• Endurance: performance VO2, MCHC and muscle oxygenation together explained 92% of the differences in time-trial power (R² = 0.92). Performance VO2 itself was explained by fibre oxidative capacity, myoglobin × capillaries per fibre and, negatively, PCSA (R² = 0.67). So a thicker muscle (hypertrophy, growth in muscle size) works against oxygen use.
• Combining both was explained by gross efficiency and performance VO2, and probably by muscle volume and fascicle length (P = 0.056 and P = 0.059). In the whole group, gross efficiency plus muscle volume explained 30%.
• Inside the muscle, fibre size and fibre oxidative capacity were inversely related (r = −0.50, a curved, hyperbolic relationship). But riders who combined fairly large fibres with a high oxidative capacity had more capillaries.
• Take-home: long fascicles (rather than a thick muscle) and many capillaries are suggested as training targets for improving sprint and endurance at the same time.

Question

Can machine learning predict a player's session RPE from many external load indicators measured with GPS and accelerometers? Session RPE (rating of perceived exertion) is the player's own score of how hard the session felt, on a 0–10 scale; it measures internal load (the body's response). External load indicators (ELIs) describe what the player physically did, such as distance, speed, accelerations and decelerations. The authors also asked which ELIs matter most for RPE in soccer, and whether a separate model per player beats one model for the whole group, i.e. whether players differ in meaningful ways.

Why it matters

Clubs track training load to improve fitness and lower injury risk. The same external load can feel very different to different players, so understanding how external load turns into internal load helps clubs manage training. Earlier studies used traditional statistics (correlation, multiple regression) on a few hand-picked ELIs. The only machine learning study was in Australian football, which has different physical demands, so its results may not hold for soccer.

Approach

38 professional outfield players from a team in the highest Dutch league were followed over two seasons (2014–15 and 2015–16). Only training sessions were used (no matches, recovery or rehabilitation sessions): 5917 sessions in total. External load came from GPS units (10 Hz, i.e. 10 measurements per second) and accelerometers (100 Hz; sensors that measure changes in speed and direction). From these the authors computed 67 ELIs on duration, distance, speed, accelerations and decelerations, PlayerLoad (an accelerometer score of total body movement) and repeated high-intensity efforts. The target (the value to predict) was session RPE on the modified Borg CR-10 scale (0–10), reported about 30 minutes after training. Three models were compared: an artificial neural network (ANN; a flexible model loosely inspired by the brain that can learn complex, non-linear patterns but is hard to interpret); LASSO (linear regression with a penalty that pushes many coefficients to exactly zero, so it keeps only the most useful ELIs); and a naive baseline that ignores all ELIs and always predicts the average RPE. Accuracy was measured on a separate test set (data the model did not learn from) with MAE (mean absolute error: the average size of the prediction error, in RPE points; lower is better). Experiment 1 trained group models on season 1 and tested them on season 2, so the test came later in time and many test players were new. Experiment 2 used the first 75% of each season to train and the last 25% to test, and compared group models (all players) with individual models (one per player). LASSO was also used to rank the ELIs by importance (how often each ELI was selected).

Findings

• Both machine learning models beat the naive baseline. Experiment 1 MAE: LASSO 0.80, ANN 1.09, baseline 1.14 RPE points. LASSO cut the error by 29.8% compared with the baseline (a small effect, d = 0.44, where d is an effect size: the difference in standard-deviation units); the ANN improved only trivially (d = 0.06).
• LASSO was more accurate than the ANN: its MAE was 26.6% lower than the ANN's in experiment 1.
• The most important ELIs included high-intensity accelerations (> 3.5 m·s−2), repeated high-intensity efforts, distance covered while decelerating hard, running distance at 12–20 km·h−1, PlayerLoad and session duration. Decelerations were the new finding: braking is eccentric work (the muscle lengthens while it contracts), which can damage muscle.
• Group models were as good as or better than individual models in both seasons (e.g. season 1 LASSO: group 0.79 vs individual 0.81 vs baseline 0.99; season 2: 0.85 vs 0.85 vs 1.11). This differs from the earlier Australian-football study. The authors explain it by far more data for a group model (> 2000 vs < 100 sessions) and a soccer squad that is more alike in load and characteristics.
• Practical message: machine learning can help experts choose which ELIs to monitor, and a group model can predict RPE for new, transferred or youth players who have little data of their own.

Question

Can one sensor on the wrist recognise three basic forearm movements while a person moves freely in daily life? The three movements are: (A) reaching for an object and bringing it back, (B) lifting a cup to the mouth, and (C) pouring or turning a key or lid (rotating the wrist). The long-term aim is a tool that counts how often a stroke (or cerebral palsy) patient uses the impaired arm for these movements outside the clinic.

Why it matters

If you can count how often the impaired arm makes these movements at home, you can follow rehabilitation from a distance: as arm function recovers, the movements should become more frequent. Earlier activity-recognition studies mostly detected whole-body activities and postures (sitting, walking, running), often in the lab. Fine arm movements in real life, with few sensors and little computing power, had received little attention.

Approach

Four healthy men and four stroke patients wore a Shimmer sensor on the back of the wrist. It is an inertial sensor: an accelerometer (measures acceleration in 3 directions, called axes) plus a gyroscope (measures rotation speed around 3 axes), recording 50 times per second (50 Hz). Training data: many repetitions of each movement under controlled conditions, labelled by the researchers (for example 480 trials per healthy person). Test data: the same person, on another day, doing a 20-step 'making a cup of tea' task freely; the researchers marked by hand where each movement started and ended. The signals were filtered to remove slow drift and fast noise. Then 10 features (summary numbers per movement) were computed for each axis, giving 30 per sensor. Examples: standard deviation (how much the signal varies), RMS (root mean square: the average size of the signal), entropy (how irregular it is), jerk (how quickly acceleration changes), number of peaks, kurtosis (how peaked) and skewness (how lopsided). The features were ranked by how well they separated the three movements and then added one at a time (sequential forward selection), keeping the set that worked best. The model was k-means clustering. k-means normally works without labels: it splits data into k groups (clusters), each with a centre (centroid); every point joins the nearest centre, the centres are recomputed, and this repeats until nothing changes. Here it was used in a supervised way (learning from examples with known answers): k was set to 3 because there were 3 known movements, each cluster got the label of the movement whose average was closest, and each new movement was labelled by its nearest cluster centre (a minimum-distance classifier). Distance was measured in two ways: Euclidean distance (the ordinary straight-line distance) or Mahalanobis distance (a distance that takes the cluster's shape into account, so a point counts as closer along a direction in which the cluster is widely spread). The feature sets were compared with 10 runs of 10-fold cross-validation on the training data (split the data into 10 parts, train on 9, test on the 10th, repeat). Each person got their own model per sensor (personalised models). The method was compared with two standard classifiers using the same features: LDA (linear discriminant analysis, which separates groups with straight boundaries) and SVM (support vector machine, which finds the boundary with the widest gap between groups).

Findings

• Healthy people: overall accuracy (share of movements labelled correctly) 61–100%, mean 88%, with accelerometer data, and 60–94%, mean 83%, with gyroscope data.
• Stroke patients: lower accuracy, 40–88% (mean 70%) with the accelerometer and 40–83% (mean 66%) with the gyroscope, because their movements were more variable and less repeatable. For example, one patient early in rehabilitation, tested on the non-dominant arm, reached only 40%.
• The number and type of best features differed per person (from 2 to 30). Healthy people shared top features (for example the accelerometer's standard deviation and RMS on the y-axis); stroke patients had little overlap.
• For stroke patients, Mahalanobis distance often worked better than Euclidean distance, because their clusters were stretched out (elongated) in some directions instead of round.
• With the same features, LDA (mean accuracy 45–53%) and SVM (mean 50–68%) did worse than the clustering method. The accelerometer and gyroscope sometimes made up for each other's weak spots.

Question

Can a machine-learning model use simple preseason tests to predict which elite youth players will get injured in the coming season? The tests covered body measurements (anthropometry), how mature the player is, motor coordination, physical fitness and years of football experience. Second question: can a similar model tell whether the injury will be an overuse injury (no single event caused it) or an acute injury (one clear event, such as a kick or a fall)?

Why it matters

Elite youth football has a high injury risk, but clubs lack the time and money for extensive screening of every player. A model built on field tests that clubs already run would help them focus prevention on the players at highest risk. Earlier studies with traditional statistics found no link between single test results and injury. Injuries usually have many causes at once, so a method that combines many variables and how they interact might reveal risk profiles that a single variable cannot.

Approach

This prospective study (players were tested first and then followed forward in time) included 734 boys from under-10 to under-15 teams (mean age 11.7 years) at seven Belgian premier-league academies, during the 2017–18 season. In August, trained testers measured: anthropometry and maturity (height, sitting height, leg length, weight, body fat, and predicted age at peak height velocity (PHV): the age at which a child grows fastest, so a higher value means a later-maturing player); motor coordination (KTK3, a children's coordination test with three tasks: jumping sideways, moving sideways and balancing backwards, plus a dribbling test); and physical fitness (jumps, sprints, the agility t-test, sit-and-reach flexibility, the Yo-Yo IR1 endurance shuttle run, curl-ups). That made 29 preseason variables. Medical staff registered each player's first injury and whether it was overuse or acute; coaches recorded training and match time. The authors built two classifiers (models that sort cases into categories): injured vs not injured, and overuse vs acute. This is supervised learning: the model learns from examples where the true answer (injured or not) is known. They used XGBoost (extreme gradient boosting): it builds many small decision trees (flowcharts of yes/no splits on the predictors) one after another, each new tree fixing the errors of the earlier ones, and combines them. The data were randomly split into 80% training data and 20% test data that the model never saw while learning. On the training data, cross-validation (repeatedly training on part of the data and checking on the rest) and grid search (trying every combination from a list of model settings, the hyperparameters, and keeping the best) tuned the model. Performance was reported as precision (of the players the model flagged as 'will be injured', the share who really got injured), recall or sensitivity (of the players who really got injured, the share the model flagged) and F1 (one score combining precision and recall; it is only high if both are high). SHAP summary plots (SHapley Additive exPlanations: a method that shows how much each predictor pushed each player's prediction up or down) showed which variables drove the predictions.

Findings

• 368 of 734 players (50%) had at least one injury. Of these first injuries, 173 were overuse and 195 acute (47% vs 53%).
• Injury model: precision, recall and F1 of 84%, 83% and 83% on the training data, and 85%, 85% and 85% on the test data (147 players). So it worked just as well on players it had not seen.
• Overuse vs acute model: 82%, 82% and 81% on the training data, and 78% for all three on the test data (74 injuries), a little lower than the injury model.
• The five most important predictors of injury (from SHAP) were a higher predicted age at PHV, taller body height, longer legs, lower body-fat percentage and standing broad jump performance. These are mostly body-size and maturity measures, which the authors link to injury rates rising with age. Better sit-and-reach flexibility raised the predicted risk slightly.
• For overuse vs acute: a lower predicted age at PHV, higher sitting height, a slower t-test and a lower moving-sideways (KTK3) score pointed towards overuse injuries.
• Conclusion: combined in a machine-learning model, simple preseason tests can flag players at high risk and the likely injury type, so academies can spend their limited resources on those players.

Question

If you group cyclists only by their own body measurements (anthropometry), using several aspects of build at once and without their discipline labels, do the groups match the disciplines they compete in (sprint, pursuit, road)? The authors also asked whether the groups differ in sprint and endurance performance, and how body build relates to performance.

Why it matters

Athletes are thought to specialise in the discipline that suits their body build. But body build is usually reported as the average of each discipline, and such group averages can hide individuals with a different build. Grouping athletes directly on their own body measures gives an unbiased test of whether build matches discipline, and helps to judge an athlete's performance in light of his build.

Approach

24 male cyclists (sprint, pursuit and road; national to Olympic level) visited the lab three times. Body measurements followed the ISAK protocol (an international standard for anthropometry). The clustering used nine variables, three for each aspect of build: body shape (the Heath–Carter somatotype: endomorphy = fatness, mesomorphy = muscularity, ectomorphy = slenderness), body size (height, weight, body surface area) and body composition (sum of eight skinfolds, body-fat %, skeletal-muscle-mass %). All variables were first turned into z-scores (value minus the mean, divided by the standard deviation), so each variable has the same scale. Then k-means clustering was run in R (Hartigan–Wong version). k-means is unsupervised learning: it gets no outcome or labels. It places k cluster centres, assigns each cyclist to the nearest centre by Euclidean (straight-line) distance, moves each centre to the mean of its cyclists, and repeats until nothing changes. The goal is to make the total squared distance between cyclists and their own cluster centre as small as possible. The number of clusters, k = 3, was chosen with the elbow criterion (plot the total within-cluster spread against k and pick the point where the curve bends), BIC (Bayesian information criterion, from the mclust package: a score that balances fit against complexity) and the NbClust package (many indices that each suggest a best k). The algorithm used 25 random starting points, and 1000 repeated runs checked that the result was stable. Afterwards, sprint performance (Wingate 1-s peak power, squat-jump power) and endurance performance (15-km time-trial power, VO2peak: the highest oxygen uptake in an all-out test), both per kg body mass, were compared between clusters (ANOVA or Kruskal–Wallis tests, which check whether groups differ) and correlated with body measures.

Findings

• Three stable clusters: mesomorphic (muscular; n = 6), short meso-ectomorphic (muscular but slim; n = 9) and tall meso-ectomorphic (n = 9).
• All six sprinters were in the mesomorphic cluster (heavier, larger girths, less lean). Pursuit and road cyclists were spread evenly over the two meso-ectomorphic clusters (4 pursuit + 5 road in each), which differed mainly in body size.
• The mesomorphic cluster had higher sprint performance (Wingate peak power, squat-jump power; p < 0.05) and lower endurance performance (time-trial power, VO2peak; p < 0.001) than both meso-ectomorphic clusters.
• Across all cyclists, better endurance went with a lean, slender build with small girths and a small frontal area (the body area facing the wind). Better sprint went with larger skinfolds and girths and a small frontal area per kg body mass.
• The idea of separate clusters for sprint, pursuit and road was only partly confirmed. The clustering separated sprint-type from endurance-type cyclists, and short from tall endurance cyclists (matching the build of all-terrain vs flat-terrain road cyclists), but not pursuit from road cyclists.

Question

Can neural networks predict lower-body joint angles and joint moments during walking directly from IMU signals? An IMU (inertial measurement unit) is a small wearable sensor; here only its accelerometer (measures acceleration) and gyroscope (measures rotation speed) were used. Joint angles describe the motion (kinematics); joint moments are the turning forces around a joint (kinetics). The goal was to make IMU-based gait analysis independent of the magnetometer (a built-in compass) and of calibration poses. A second question: how does a standard feedforward network compare with a recurrent LSTM network on this task?

Why it matters

Gait analysis in the lab uses cameras and force plates (plates in the floor that measure the force of each step). This is expensive and only works in a small recording space. IMUs work anywhere, but they measure joint angles only indirectly, depend on calibration that easily goes wrong and on magnetometers that are easily disturbed by local changes in the magnetic field, and they cannot measure joint moments at all. If a network can turn raw IMU signals into gold-standard joint angles and moments (gold standard = the best available measurement), full gait analysis becomes possible in clinics and in the field.

Approach

The authors reused a database of walking trials on level ground: people without movement disorders plus patients with a knee replacement, walking at their own chosen speed (0.8–2.0 m/s). The trials had been recorded with optical motion capture (cameras tracking reflective markers on the skin) and force plates. Joint angles were computed from the markers. Joint moments were computed with inverse dynamics (working backwards from the measured motion and floor forces to the forces at each joint) in the AnyBody software, and scaled to body weight and height. These angles and moments were the targets (the known answers). From the marker paths the authors calculated what virtual IMUs on the body would have measured (acceleration and rotation speed, no magnetometer). This gave model inputs that match the gold-standard targets exactly. They trained two kinds of artificial neural network (a model made of layers of connected simple units that learns a flexible input-to-output function from examples): a feedforward network (FFNN), which sees the whole gait cycle or stance phase at once, and an LSTM (long short-term memory) network, a recurrent network that reads the signal step by step and keeps an internal memory of what came before. Training used standard settings: mean squared error as the quantity to minimise, the Adam training algorithm, early stopping (stop training when the error on validation data stops improving) and dropout (randomly switching off units during training); both reduce overfitting (learning noise in the training data so the model fails on new data). The data were split into training, validation and test sets by participant, so no person appeared in more than one set. They started from a baseline (IMU data, right steps only, trained 10 times) and tested changes: adding all or some body measurements, using both left and right steps, and data augmentation (creating extra training examples by randomly rotating the virtual sensor). A change was kept only if it beat the baseline by more than two standard deviations. Accuracy was reported as the correlation coefficient r (how closely the predicted curve follows the shape of the true curve) plus RMSE (root mean square error: typical size of the error; in degrees for angles) or nRMSE (RMSE divided by the range of the curve, in %, for moments). They reported both because r alone ignores an offset: a curve shifted up by 5° can still have r = 1.

Findings

• Both networks predicted joint angles and moments well. The abstract reports mean correlations above 0.98 in the sagittal plane (forward-backward movement seen from the side) and above 0.80 in the smaller planes (frontal: sideways; transverse: rotation). The Discussion reports RMSE below 4° (angles) and nRMSE below 25% (moments) for all joints and planes. Caveat: a few single values are lower, for example knee frontal-plane angle r = 0.79 (FFNN) and 0.68 (LSTM) in Table 4.
• The feedforward network beat the LSTM for both angles and moments in most joints and planes (baseline joint-angle RMSE 1.60° vs 2.15°; baseline moment nRMSE 12.2% vs 14.8%). The authors' explanation: the FFNN sees all time steps at once, while the LSTM can only carry information forward from earlier time steps.
• Sagittal-plane motion was predicted best. For joint angles the transverse plane was worst; for joint moments the knee rotation moment and the frontal-plane ankle moment were worst.
• For joint moments, using both left and right steps plus data augmentation (3 times as much data) improved the models. For joint angles, only adding some body measurements (height, weight, segment lengths, pelvis width) helped.
• Conclusion: the FFNN suits analysis of fully recorded data afterwards; the LSTM is the more natural choice for real-time use and recordings of any length. A convolutional network with a small delay might combine both strengths.

Question

How is big health data used and analysed to create value for patients, analysts and managers in healthcare? The authors use the resource-based view, a business theory that says an organisation gains value by combining its resources (here: data) with its capabilities (here: analysis). They map which data types are used, which analysis techniques are applied, which values are created, which platforms and tools handle big health data, and which future directions are proposed.

Why it matters

Healthcare produces huge amounts of data: electronic health records (EHR: the digital patient file), lab systems, medical images, home sensors, billing data and social media posts. Big data are defined by the 4 Vs: volume (very large amounts), velocity (data arrive fast, often in real time), variety (many formats: structured, unstructured and semi-structured) and veracity (whether the data are reliable). The number of papers on health analytics is growing fast, so doctors, policy makers and researchers need an organised overview that links technical progress to the value it creates.

Approach

A systematic literature review: a review that finds all papers on a topic with a fixed, repeatable search and fixed rules for which papers count, done in two steps (following Tranfield et al., 2003, and Dubey et al., 2017). Step 1: they searched two databases of scientific papers, Web of Science and Scopus (2000–2016), for 'business intelligence', 'analytics' and 'big data' combined with health, medical and clinical terms. They kept only English-language articles and reviews about both health and big data analytics: 6817 hits → 3241 after removing duplicates → 1877 after reading titles and abstracts → 804 included after reading the full texts. Step 2: content analysis, meaning they read each paper and coded (labelled) it, using the text-coding software NVivo, by data type, technique, value created and future direction. A paper could get several labels. Then they counted how often each label occurred and made cross-tables, for example technique by data type and technique by value.

Findings

• Data types: clinical data (for example EHRs and medical images) dominate, in 69.9% of papers (562 of 804). Next come patient behavior and sentiment data (from wearable sensors and social media; 16.5%), administrative and cost data (7.3%) and pharmaceutical research data (4.7%).
• Techniques: modelling (42.8%) and machine learning (40.7%) were used most, followed by data mining (finding patterns in large data sets; 24.9%), visualisation (19%) and statistics (16.4%). Machine learning (algorithms that learn patterns from data instead of following fixed rules) was the most-used technique across almost all values and data types, so the authors give it its own section.
• Values created: the most frequent was better diagnosis for more personalised healthcare (35.6%), then supporting or replacing professionals' decisions with automated algorithms (25.6%) and new business models, products and services (24.5%). The values are grouped by who benefits: patients (P), analysts (A) or management (M).
• Platforms: Apache Hadoop is the most-used computing platform. It stores data spread over many computers (HDFS) and processes them in parallel (MapReduce). A main difficulty is that most health data are unstructured (for example the free text of doctors' notes, not neat tables), which makes them hard to process by computer.
• Future directions are mostly technological; the most frequent was developing the presented approach further (16.5%). For machine learning in particular they name: unsupervised learning (finding groups without known labels) to describe subtypes of complex diseases more precisely, automated risk prediction to guide care, reinforcement learning (learning by trial and reward) to support care providers, the tension between accuracy and interpretability (how easy it is to understand why a model gives its answer), and training clinicians to judge ML-based tools.

Question

Can you estimate how fast someone runs 10 km using only data collected in daily life, without any lab tests? The data were morning heart measurements from the HRV4Training phone app, details the users typed in themselves, and running workouts copied in automatically from other apps. The authors also asked which groups of predictors matter most, and how many workouts you need for a good estimate.

Why it matters

A good estimate of running performance helps runners pace their races and fit their training plan to their ability, which may also lower injury risk. Earlier studies had three problems. They used small groups of similar people (for example only men). They needed lab measures such as VO2max (the most oxygen the body can use per minute) or lactate threshold, which ordinary runners cannot measure. And they often did not test their models on new people. Data from apps and wearables offer much larger, more realistic samples.

Approach

Users of the HRV4Training app agreed to share their data. Each morning they measured resting heart rate (HR) and heart rate variability (HRV: how much the time between heartbeats varies; higher HRV usually means a more rested body) with the phone camera, which had been checked against an ECG (the clinical heart monitor), or with a chest strap. Their running app (Strava or TrainingPeaks) was linked to HRV4Training through an API (a connection that lets one program fetch data from another), so every workout with GPS and heart rate came in automatically. In 2016–2017, 2113 users qualified: they trained with a heart-rate monitor, had linked workouts, had run at least one 10 km and had at least one month of morning measurements. The outcome (the 'true answer' the model learns from) was each user's fastest 10 km found in their own workouts. The predictors came from the 3 months before that run and were added in groups, one step at a time: (1) anthropometrics: body data such as BMI (weight relative to height), age and gender; (2) resting physiology: morning HR and rMSSD (the standard HRV number); (3) training volume and speed: how far and how fast they ran; (4) training physiology: speed divided by heart rate during runs, a fitness marker; (5) training polarization: how much of the training was clearly easy or clearly hard instead of moderate, measured as the share of workouts more than 5% faster or slower than the user's own average, and the share at moderate heart rate; (6) previous performance. The authors measured polarization relative to each user's own average, not with maximum, minimum or range, because single wrong readings from everyday sensors would distort those. The model was multiple linear regression (predicts a number as a weighted sum of the predictors). They chose it because its coefficients (the weights) are easy to read: each weight shows how strongly a predictor goes with a faster or slower time, and in which direction. This is supervised learning: the model learns from examples where the true answer is known. They tested it with 10-fold cross-validation: split the users into 10 parts, train on 9 parts and test on the left-out part, and repeat 10 times. Only users with at least 20 workouts were used to train, but all users were tested. Accuracy was reported as RMSE (root mean square error: the typical size of the error, here in minutes), mean percentage error, r (correlation between estimated and true times) and R² (the share of the differences between runners that the model explains, from 0 to 1).

Findings

• The error went down with every group of predictors added: RMSE 6.27 min with body data only, 6.07 with morning HR and HRV, 4.04 with training volume and speed, 3.96 with training physiology, 3.64 with training polarization and 2.68 min with previous performance (Table II). The biggest single drop came from adding training volume and speed.
• The best model (all predictors, including previous performance) had an RMSE of 2.6 min in the abstract (2.68 min in Table II), about 4% error, r = 0.93 and R² = 0.87. That is 58% better than using body data alone.
• About 15 workouts were enough to compute useful training predictors and get the error close to its minimum (Fig. 3).
• Direction of the coefficients: younger age, lower BMI, lower resting HR, higher HRV, longer and faster workouts, more workouts clearly faster or slower than one's own average, and less time at moderate heart rate all went with a faster 10 km. In short, more polarized training (mostly easy plus some hard, little moderate) went with better performance.
• These links agree with earlier small lab studies, but here they were found without any lab testing. So recreational runners could use such a model to estimate their performance and adjust their training.

Question

Where does wearable sensing stand after about ten years of fast commercial growth, and what comes next? The authors look at what wearables measure now, the shift from tracking behavior to tracking health, large-scale research with wearable data, the main open problems and the likely next steps.

Why it matters

Wearables are now mass-market products: the Apple Watch has sold more than 100 million units (up from 30 million in 2017) and Google bought Fitbit. They now make observational studies (studies that only watch people, without changing anything) with thousands of participants possible, including tracking infections during COVID-19. At the same time, people started to question how accurate and how biased (systematically wrong for some groups) the sensors are. Researchers therefore need a clear view of what wearable data can and cannot support.

Approach

This is a perspective column, not a study with new data: the authors summarize recent developments and a few chosen studies. They describe what is measured now: the standard phone sensors (GPS for location, accelerometers and gyroscopes for movement, the camera for heart measurements), blood pressure from PPG (photoplethysmography: an optical sensor that shines light into the skin and measures changes in blood flow) on the wrist or ear, earables (in-ear devices such as AirPods), and glucose measured in interstitial fluid (the fluid between cells under the skin) instead of blood. Measuring glucose from sweat has so far failed. They describe the shift from tracking behavior to tracking physiology (the body's own signals), and a three-step route to new knowledge, explained with a Fitbit heart-rate example and with the COVID-19 recovery data of Radin et al. (2021). Then they discuss the open problems (accurate data, wearability and access to the body, energy use) and look ahead (new sensors, devices that act as well as measure, and forecasting with machine learning).

Findings

• The focus has moved from absolute measures of behavior (steps, energy expenditure) to relative changes in physiology: how a person's resting heart rate (HR), heart rate variability (HRV: how much the time between heartbeats varies) or blood glucose change compared with their own normal level, in response to stressors such as sickness, alcohol, exercise, diet and the menstrual cycle. Activity and location are still useful as context.
• Three steps to new knowledge: (1) the device must measure accurately, shown by validation (comparing it with a reference system) and certification (official approval as a medical device, such as FDA approval or a CE mark); (2) it must reproduce known results from smaller clinical studies, for example the short-term heart-rate response to exercise or sickness; (3) only then can use at large scale reveal new relationships, for example COVID-19, long COVID and the return to one's normal physiology (Radin et al.).
• Possibly the biggest challenge is the noise of everyday, unsupervised settings: sensors can be worn wrongly or fail (for example optical HR during exercise), and devices rarely report signal quality or how confident they are in a value. Skin color, motion artefacts (errors caused by movement) and sensor position also reduce accuracy: a watch sits on top of the wrist, but the large blood vessels run underneath, so the sensor must rely on tiny vessels (capillaries).
• Other challenges: a device in one spot (the wrist) can measure only a limited set of things, while sensors built into clothing (e-textiles) are hard to develop, test and manufacture. Energy use and battery power also remain problems.
• Next steps: new types of sensors, with more attention to bias and inclusion; closed-loop systems that act as well as measure (for example a glucose monitor that steers an insulin pump, or vagal nerve stimulation, a mild electrical stimulation at the ear meant to reduce stress); and machine-learning models that forecast an individual's response (for example tomorrow's HRV if you skip a glass of wine today) instead of only describing past changes.

Question

After COVID-19, how long does it take for resting heart rate (a body signal) and for sleep and daily steps (behavior) to return to the person's own pre-illness level, called their baseline? How much does this recovery differ between people? The comparison group was people with similar symptoms who tested negative for COVID-19.

Why it matters

Long-term problems after COVID-19, such as autonomic dysfunction (problems with the part of the nervous system that automatically controls heart rate and other body functions) and heart damage, had been reported for up to 6 months but had not been measured in numbers. Wearables (sensors worn on the body, such as smartwatches and fitness trackers) record body signals and behavior continuously, starting while people are still healthy. So they capture a person's healthy baseline, the illness and the recovery, which a normal clinical study rarely does.

Approach

DETECT is a remote, app-based cohort study (a study that follows the same group of people over time) that enrolled 37,146 adults across the US between March 2020 and January 2021 and collected their wearable data. It is observational: the researchers only watched, they did not change anything. This analysis used the 875 participants who reported symptoms of an acute respiratory illness (a sudden infection of the airways) and had a COVID-19 swab test: 234 tested positive and 641 tested negative (the comparison group). The swab test was the reference: the verified answer to 'who had COVID-19'. For each person, daily resting heart rate (RHR) was expressed as the difference from their own baseline: daily RHR minus their average RHR before illness. Sleep and daily step count were also followed, from 7 days before to 133 days after symptoms started. The COVID-positive people were then split into three groups by how far their RHR stayed above baseline 28–56 days after symptoms started (less than 1, 1–5, or more than 5 beats per minute, bpm). Symptoms in these groups were compared with χ² tests (tests whether how often something occurs differs between groups) and age with ANOVA (a test that compares the means of three or more groups). The analysis is descriptive: average curves with 95% confidence intervals (the range that very likely contains the true average). There is no prediction model.

Findings

• People with COVID-19 took longer than symptomatic people who tested negative to return to their baseline for resting heart rate, sleep and activity.
• The difference was largest for RHR: first a short drop below baseline (bradycardia: a slower heart rate than normal), then a long rise above baseline (tachycardia: a faster heart rate than normal) that on average did not return to baseline until 79 days after symptoms started. Step count and sleep returned sooner, after 32 and 24 days.
• A small group of 32 COVID-positive people (13.7%) kept an RHR more than 5 bpm above baseline that had not returned to normal after more than 133 days.
• In the first days of illness this group reported more cough (84.4%), body ache (62.5%) and shortness of breath (28.1%) than the other groups. This suggests that early symptoms and a larger first rise in RHR may go together with a longer recovery.
• Overall, COVID-19 affected the body for about 2–3 months on average, with large differences in recovery. This may reflect problems in the autonomic nervous system or ongoing inflammation.