REF-1439 · Rothschild et al., 2025 · Sports Medicine · Paper · PubMed · Free full text

Can I add bike TSS and run TSS into one fitness number?

Not as like units: cycling-built models over-predicted runners' session carbohydrate (power TSS by 35 g, other scores 26-50 g) until recalibrated per sport. Energy, not adaptation.

Main caveat: Outcome is fuel and energy cost of single sessions, estimated by a new method checked only against published data.
limited evidenceMeasured on trained athletes

Rothschild and colleagues (High Performance Sport New Zealand, Auckland University of Technology) asked whether common training-load measures track the carbohydrate and energy cost of a session, and whether models built on them can predict that cost.

In the primary study 15 cyclists (5 women; VO2max 51.9 ± 7.2 mL/kg/min) rode four sessions in randomised, counter-balanced order — 90 and 30 min at 90% of VT1 power, and two interval sessions — and carbohydrate use and energy expenditure were estimated second by second from gas exchange (aerobic) and from net lactate accumulation (anaerobic).

Six load metrics — total work, session RPE load, Lucia TRIMP from power and from heart rate, TSS from power and TSS from heart rate — correlated very strongly with both outcomes within subjects (r = 0.71-0.98 in the Results; 0.76-0.96 in the abstract).

Prediction models adding fitness, sex and prior-day load explained 77-96% of variance, with cross-validated mean absolute errors of 17.6-21.2 g carbohydrate and 58.5-81.7 kcal. TSS was the only metric that gave the same value to the 90-min easy ride and the short-interval session, which cost the same carbohydrate.

In the validation study 59 new athletes did one session — 30 min at 95% of VT1 then 6 × 3 min hard — in their own sport. Predictions were best in cycling; in runners and kayakers most models over-predicted, and the authors call the between-sport differences "highly consistent".

Errors ranged 15-45% for carbohydrate and 9-31% for energy, and a simple linear recalibration per sport cut them to 8-18% and 6-12%. Our reading: the only study in the base that applies the same load scores across sports against a measured outcome.

Within a sport the scores track cost well; across sports the same relationship did not hold until recalibrated — so a ride's score and a run's score are not one currency, at least for energy and carbohydrate. It says nothing directly about whether equal scores bring equal adaptation or fatigue.

Protocol

Primary study: n = 15 cyclists (10 men, 5 women), five lab visits: a graded test (VT1, VT2, VO2max, peak power), then four sessions in randomised, counter-balanced order after an overnight fast and a standard snack (27 g carbohydrate) 45 min before — LIT-long (90 min at 90% VT1 power), LIT-short (30 min), HIIT-long (15 min warm-up, 2 × 5 × 3 min at 110% VT2 power), HIIT-short (15 min, 2 × 10 × 30 s).
Validation study: n = 59 (21 cyclists, 20 runners, 18 kayakers; 41 men, 18 women), one session each: 30 min at 95% of VT1, 5 min recovery, 6 × 3 min with 2 min rest (cycling at 110% VT2 power in a cadence-dependent mode, running at 107% VT2 speed, kayak self-paced); running power from a Stryd foot pod, kayak power from the ergometer.
Outcome: carbohydrate use and energy expenditure, aerobic part from VO2 and RER second by second (Elia and Livesey energy equivalents, glucose-glycogen split scaled to %VO2max), anaerobic part from net lactate accumulation × body mass × 3 mL O2 per mmol/L; checked against published tracer and glycogen studies (supplement, not read).
Load metrics: total work (kJ); session RPE × minutes (CR100 scale ÷ 10); Lucia TRIMP (zones at VT1 and VT2) from power and from heart rate; TSS from power with VT2 power as FTP; heart-rate TSS with VT2 heart rate as threshold.
Analysis: repeated-measures correlation; generalised estimating equations with model selection by corrected AIC and participant-level leave-one-out cross-validation; validation by R², mean absolute error, calibration-in-the-large and calibration slope; recalibration by regressing measured on predicted values per sport.
Funding and interests (as printed): no sources of funding for the article (open access funded through CAUL); Rothschild is on the scientific advisory board of Applied Behavior Systems Ltd; the other authors declare no relevant conflicts. Data and R code on OSF.

Key Results

Primary study (cyclists)Result
Within-subject correlation of each load metric with carbohydrate use and energy expenditurer = 0.71-0.98 (abstract: 0.76-0.96)
Model fit, as the abstract and Results state itR² 0.77-0.96; MAE carbohydrate 16-21 g (10-14%), energy 53-82 kcal (7-11%); these ranges mix Table 3's in-sample values (R² 0.88-0.96; MAE 16.0-19.1 g, 52.6-73.1 kcal) with its cross-validated ones (cv-R² 0.77-0.93; cv-MAE up to 21.2 g and 81.7 kcal)
Model fit, cross-validated (Table 3)carbohydrate MAE 17.6-21.2 g; energy 58.5-81.7 kcal
Day-to-day typical error, easy cyclingheart rate 2.9 beats/min; VO2 0.08 L/min; carbohydrate oxidation 0.21 g/min; RPE 4.0 (0-100)

Validation in new athletes, by sport (Table 4; MAE before calibration / calibration-in-the-large, measured minus predicted / MAE after sport-specific calibration):

MetricCyclists, carbohydrate (g)Runners, carbohydrate (g)Kayakers, carbohydrate (g)Cyclists, energy (kcal)Runners, energy (kcal)Kayakers, energy (kcal)
TSS (power)18.3 / −2.3 / 17.836.9 / −35.4 / 18.431.9 / −27.2 / 11.268.1 / 2.2 / 65.3113.0 / −103.2 / 54.191.1 / −46.7 / 45.1
TSS (heart rate)23.5 / −11.2 / 19.329.5 / −25.5 / 15.028.9 / −25.2 / 8.580.3 / −20.7 / 72.481.4 / −38.1 / 55.181.9 / −43.5 / 43.4
Session RPE load21.9 / −0.8 / 20.332.0 / −28.4 / 18.629.9 / −20.2 / 15.079.9 / 8.7 / 70.691.6 / −48.5 / 58.4106.3 / −43.1 / 59.7
Total work (kJ)16.7 / −7.4 / 15.050.4 / −50.0 / 15.716.1 / 14.2 / 11.753.0 / −18.5 / 48.3178.3 / −178.3 / 39.5181.8 / 181.8 / 36.0
Across all six metrics: raw MAE 16.1-50.4 g (15-45%) for carbohydrate and 53.0-181.8 kcal (9-31%) for energy; after calibration 8.5-20.3 g (8-18%) and 36.0-72.4 kcal (6-12%), with the calibration fitted and scored on the same validation athletes, so these errors are optimistic (no independent test of the calibrated models). In runners, every model over-predicted carbohydrate (calibration-in-the-large −25.5 to −50.0 g).
TSS and session type: only TSS rated the 90-min easy ride and the short-interval session alike, as their carbohydrate cost was alike; session RPE load rated the interval session higher although total work was higher in the long ride.
Gross efficiency (not shown in tables): 20.0% cycling, 10.1% kayaking, 26.2% running (running power from the Stryd pod).
Numbers that do not add up: the abstract gives within-subject correlations of 0.76-0.96 and the Results and Discussion 0.71-0.98; the Discussion's "MAE 17–24 g" for cycling does not match Table 3 (16.0-19.1 g in-sample, 17.6-21.2 g cross-validated).

Mechanisms

Measured: within cyclists, all six scores rose with the carbohydrate and energy cost of the session; across sports the misses were systematic by sport and largely removed by a linear calibration.
Authors' interpretation: differences inherent to each sport and its measuring devices, not individual noise, cause most of the error; the running power meter seems to overstate power relative to cycling, and gross efficiency differs by mode. TSS's quadratic weighting of intensity may be a "feature" for carbohydrate, because anaerobic glycolysis uses far more carbohydrate per ATP.
Inference: a run and a ride with the same score did not carry the same carbohydrate or energy cost here, so a fitness line that sums them adds unlike units for energy; for heart-rate TSS the running miss was smaller than for power TSS or total work, but still present. Mechanical and tissue cost, which differs most between running and cycling (REF-235, PHYS-22), is outside every one of these scores.

Practical Application

Can I add bike TSS and run TSS into one fitness number? Not as like units: cycling-built models over-predicted runners' session carbohydrate (power TSS by 35 g, other scores 26-50 g) until recalibrated per sport. Energy, not adaptation.
Keep a load line per sport: read bike and run scores separately and use session RPE as the common internal measure within each sport (PHYS-45, SYNTH-39); a summed CTL is not running fitness after weeks of riding. Not from this paper alone; the base's own practice, supported by this measurement.
Within cycling the scores are good proxies for fuel cost: all six tracked carbohydrate and energy closely in cyclists (runners and kayakers were tested only in the validation, where the cycling-built models needed recalibration); TSS best reflected the extra carbohydrate cost of hard intervals.
What it does not show: that equal scores across sports mean equal adaptation or fatigue (not tested), anything beyond 90 min or outdoors, or swimming.

Limitations

Outcome is fuel and energy cost of single sessions, estimated by a new method checked only against published data; adaptation, fatigue and injury were not measured.
Models built on 15 cyclists; the authors note the sample is below that needed for precise residual estimates; validation arms of 18-21 per sport, the kayak arm just below its target for one model.
Device dependence: running power from one foot pod; FTP estimated as VT2 power; HR-TSS with VT2 heart rate.
Laboratory conditions, standard snack, 30-90 min; may not apply to fasted, long or outdoor sessions.
Reporting slips in correlation and MAE ranges, noted under Key Results; the supplement was not read.
Interests: first author on the scientific advisory board of Applied Behavior Systems Ltd; no funding for the work.
How this entry was checked and changed
Why this level

a single-laboratory validation study (the rubric's B for test validation), about 20 athletes per sport in the validation; the outcome is the carbohydrate and energy cost of one session estimated by a new aerobic-plus-anaerobic method, not training adaptation or fatigue

Full text (checked 2026-10-29)
The primary study included 15 cyclists (five female); the validation study included 21 cyclists (seven female), 20 runners (six female), and 18 kayakers (five female) (full text)
Very-large to near-perfect within-subject correlations (r = 0.76–0.96) were evident between all training load metrics and both CHO utilization and EE (abstract); very large to near-perfect associations (r = 0.71–0.98) (Discussion) (full text)
In the validation study, the mean absolute error ranged from 16–50 g [15–45%] for CHO models to 53–182 kcal [9–31%] for EE models; the calibrated mean absolute error ranged from 9–20 g [8–18%] to 36–72 kcal [6–12%] (full text)
Overall, predictions for cycling were better than running and kayaking, and the between-sport differences were highly consistent (full text)
Carbohydrate utilization (g), run, TSS: MAE raw 36.9, calibration-in-the-large −35.4, MAE calibrated 18.4 (full text)
TSS was the only measure of training load to accurately reflect similar total carbohydrate utilization between the longer duration low-intensity session and the shorter duration high-intensity session (full text)
Mean gross efficiency values in our study were 20.0% for cycling, 10.1% for kayaking, and 26.2% for running (full text)
The typical error of measurement of day-to-day variation during low-intensity cycling was 2.9 beats/min for HR and 4.0 arbitrary units (0–100 AU scale) for RPE (full text)
Jeffrey A. Rothschild is on the scientific advisory board of Applied Behavior Systems Ltd.; no sources of funding were used to assist in the preparation of this article (full text)
Can I add bike TSS and run TSS into one fitness number? — BLOCKS Science · Blocks