When a hospital adopts a new surgical technology such as a robotic platform, patients operated on early in the adoption period face extra risk. The surgeon is still learning, and so is the team that positions the patient, docks the robot and assists at the bedside. These two learning processes run on different clocks. In one robotic colorectal series, the console surgeon's efficiency stabilized after 31 cases, while the team's setup and docking time needed 60 cases to mature (Learning curve and surgical outcome of robotic assisted colorectal surgery with ERAS program). Both counts rise together, case after case. So when a patient has a complication at case 45, the data cannot tell, without more structure, whether the excess risk belongs to the surgeon, to the team, or to nobody. Environmental scientists call this a source attribution problem: a single measured outcome mixes contributions from several sources that move together. They solve it by finding variation that moves one source and not the other. Wind direction is the classic example, because it shifts a power plant's plume relative to a monitor without changing anything else (Quantifying the impact of air pollution from coal-fired electricity generation on crop productivity in India).
We propose to carry that logic into time-to-event modeling. We will build a proportional hazards model, meaning a survival regression in which each covariate scales a shared baseline risk of an event over follow-up time. In this model the patient's hazard of complication, readmission or reoperation depends on two separate, shape-constrained learning curves: one in the surgeon's cumulative case count and one in the assisting team's cumulative shared case count. The two curves are identified by the variation that hospitals generate as a matter of routine. Staff rotate between rooms, several surgeons share one team, assistants leave and new ones arrive. Each of these events resets team experience without touching surgeon experience, much as wind direction does for the plume. The result is a learning-curve estimate that answers an operational question existing methods cannot: which group's training would most reduce patient risk.
Core Claim
Our claim is that surgeon and team learning effects on censored patient outcomes can be separated, with honest uncertainty, if and only if the case log contains crossed experience variation. We will show this in two ways. A two-curve hazard model will recover both curves in calibrated simulations where pooled models cannot. On real registries, it will predict the transient rise in risk that follows team turnover, which a single pooled experience curve cannot predict by construction. Where crossed variation is absent, we expect the method to say so through wide intervals rather than invent an attribution. That honesty is part of the contribution.
What Exists and What Does Not
The learning-curve literature measures thresholds with cumulative sum charts, or CUSUM: running totals of each case's deviation from a target operative time or complication rate, plotted against case number so that the inflection point marks the end of the learning phase. Examples include the 31/60-case split above, about 20 cases for en bloc bladder tumour resection (En bloc resection of bladder tumour: the rebirth of past through reminiscence), about 35 cases for heads-up display microsurgery (Asani et al.), and 40 to 50 cases for complication stabilization in endoscopic skull-base surgery (Gupta et al.). These studies motivate the proposal: their thresholds are evidence that risk is concentrated early. None of them models censored patient outcomes, and only Lin et al. separates surgeon from team, and then only for operating-room minutes. Duration is also a weak proxy for safety (Bibliometric analysis of research trends and hotspots on robot-assisted thyroid surgery).
The survival literature runs the other way. It models censored outcomes well but treats the operation as a fixed event. Krajewski et al. used propensity matching, change-point regression and Cox models across 13 centers to show that delayed adjuvant BCG immunotherapy raises the hazard of recurrence (HR 1.62) and progression (HR 3.34) (Delaying BCG immunotherapy onset after transurethral resection of non-muscle-invasive bladder cancer is associated with adverse survival outcomes). They never record where the resecting surgeon sat on a learning curve, even though they note that residual tumor at re-resection weakened the timing effect. That is exactly the kind of mixed-source signal we target. DeepSurv replaces the Cox linear predictor with a neural network and modestly improves discrimination on static tabular data (Deep Survival: A Deep Cox Proportional Hazards Network). It supplies a flexible comparison model, but flexibility does not resolve collinearity between two experience counts.
Two papers constrain the design. Hendricks et al. found that nominal novice/intermediate/expert labels correlate minimally and inconsistently with rated skill on JIGSAWS, a benchmark of robotic training-task recordings, and that robot-side kinematics track skill better (Exploring the Limitations and Implications of the JIGSAWS Dataset for Robot-Assisted Surgery). Case counts are therefore a noisy index of skill, and our curves must be read as effects of exposure, not of competence. Li et al. found null hazard ratios (PM2.5 HR 1.01, p=0.599) when entering correlated pollutants one at a time into time-dependent Cox models (Preliminary investigation of the association between air pollution exposure and childhood asthma hospitalizations from 2015 to 2018 in East China). This is a warning that collinear exposures entered naively dilute or hide effects, the failure mode we want to avoid.
The Model
For patient $i$ operated on as surgeon $s$'s $n_{s}$-th case, with a team whose members have shared on average $m_{i}$ prior cases on this platform, the hazard at follow-up time $t$ is
$$h_i(t) = h_0(t)\,\exp\big(\beta^\top x_i + f_S(n_{s}) + f_T(m_{i}) + u_s\big).$$
Here $x_i$ are patient risk factors, $u_s$ is a surgeon-level random effect (a "frailty," capturing persistent differences between surgeons), and $f_S, f_T$ are monotone non-increasing penalized splines anchored at zero for large counts. Each therefore reads as the excess log-hazard of inexperience.
Two refinements address known problems. First, learning effects plausibly act mostly in the early postoperative window: a technical error causes an early leak, not a late one. A static Cox model with a case-count term would violate proportional hazards in exactly this way. We therefore allow $f_S$ and $f_T$ to be multiplied by a simple decaying function of $t$, and we test proportionality explicitly. Second, the source-apportionment receptor models of atmospheric science use chemical fingerprints to tag sources. We borrow that idea in a modest form: recorded intermediate times serve as auxiliary signatures, with docking/setup time tagging the team and console time tagging the surgeon. They enter as a jointly modeled side outcome that shares the $f_S$ and $f_T$ curves, which lends strength to the decomposition.
Identification comes from team composition that varies independently of surgeon count. We quantify the available variation before fitting, through the conditional variance of $m_i$ given $n_s$. We report it as a diagnostic, analogous to a first-stage strength check for an instrument.
Research Program
The work has three stages. First, a simulation study calibrated to published structure: surgeon plateau near 31 cases, team plateau near 60 (Lin et al.), and realistic censoring and rotation rates. We vary the amount of crossed variation from none to heavy and measure recovery of both curves and of their plateau points. Second, application to multi-surgeon robotic surgery registries that log staff rosters, operating-room timestamps and 90-day outcomes. The cohort types described in Lin et al. and in the multicenter bladder registries (Krajewski et al.; the ERBT registries in En bloc resection of bladder tumour) have this shape. These are not public, and access through collaborating centers is a real dependency. Third, a check that the constrained additive model gives up little predictive accuracy relative to flexible survival networks. Sufficiently large clinical registries are not in hand, so this check initially runs on the public survival benchmarks used by DeepSurv.
The primary outcomes are time to first complication within 90 days, and time to discharge with in-hospital death as a competing event. Secondary outcomes are reoperation and, in oncologic cohorts, recurrence-free survival.
Baselines
The naive reference is a Cox model with patient covariates and surgeon frailty but no experience terms. It tells us how much learning effects matter at all. Any proposed model that does not improve held-out partial likelihood over it has found nothing.
The conventional clinical practice baseline follows Lin et al.: CUSUM thresholds are estimated separately on console time and setup time, cases are dichotomized into early and late phases at those thresholds, and the phases are entered as indicators in a Cox model. This baseline is fair to clinical practice because it uses the same two-source logic, only discretized. It shares our training data and our patient covariates.
The strongest existing competitor is a single pooled learning curve: a Cox model with a penalized spline in surgeon case count plus surgeon frailty. This is the standard modern way to put a learning curve into a hazard model. On ordinary held-out prediction it will likely match our model, because surgeon and team counts are correlated. It gets the same spline basis, the same smoothing-selection procedure (likelihood-based, no extra tuning budget), and the same data.
The flexible competitor is DeepSurv given both counts and all patient covariates, tuned with the same number of hyperparameter trials as our model's smoothing search. On METABRIC, DeepSurv reports a concordance index (C-index, the probability that the model ranks a random pair of patients correctly by risk) of 0.643, against 0.630 for linear Cox and 0.624 for random survival forests (Deep Survival). On METABRIC we treat a C-index within about 0.01 of 0.643 as "no meaningful accuracy cost." No reported result was found on Lacuna for any learning-curve survival model on a surgical registry, so our real-data comparisons will be internal.
Two ablations isolate the key ingredient. The first refits our model on the subset of cases where team composition never changed, removing crossed variation. If the curves are truly identified by rotation, their intervals should widen sharply and the split between them should become arbitrary. The second removes the auxiliary timing signatures, to measure how much they add beyond rotation alone.
Beating the strongest baseline means three things. In simulation, lower integrated error on both component curves and nominal interval coverage for the team plateau point, where the pooled model is structurally biased. On real data, held-out partial likelihood no worse than the pooled model. Most importantly, significantly better calibration on the specific cases that follow a team member's arrival.
Why not just…?
Why not just run CUSUM separately on console and setup time, as Lin et al. did? That decomposition attributes operating-room minutes, not patient harm. Minutes and safety can diverge: robotic cases took 60 to 90 minutes longer yet had shorter stays and fewer minor complications in the same series. CUSUM also has no way to handle censored follow-up. Our discretized-phase baseline is precisely this approach carried into a Cox model, so the experiments test it directly rather than dismiss it.
Why not just put a spline in surgeon case count with a surgeon frailty? This is the real threat. If a hospital's teams never rotate, our model collapses to it and should. The difference is only visible where surgeon and team experience come apart. The decisive test is therefore prediction after team turnover: a pooled surgeon curve predicts no change in risk when a new assistant joins, while the two-curve model predicts a transient rise. If the pooled model is as well calibrated on those cases, team learning does not matter for outcomes, and we will report that as a finding.
Why not just give DeepSurv both counts and let it learn the interaction? A flexible network fitted to nearly collinear inputs will split their effect in whatever way the optimizer happens to find. It will not produce a decomposition with uncertainty, and it has no notion of monotonic learning. The simulation measures this directly by comparing recovered component curves, computed for DeepSurv by partial dependence, against the truth across levels of crossed variation.
Why not just use crossed random effects for surgeon and each team member? Static effects capture who is better, not how risk changes with shared experience. We include the surgeon effect for that reason. Adding member-level static effects is a sensible extension, and we will test it as a sensitivity analysis.
Decisive Checks
Three results settle the question. In simulation, the two-curve model should recover the team plateau within its interval at moderate rotation rates, while the pooled model misplaces the plateau and misattributes team risk to the surgeon. In the no-rotation ablation, attribution should honestly fail. On real registries, staff-arrival events form a natural experiment. The model fitted on data before each arrival should predict an elevated post-arrival hazard that the pooled model misses, measured by calibration of observed against expected 90-day events in the 20 cases after each arrival. If that prediction fails, the claim that team learning carries patient risk is wrong for that setting, and the method has still done its job by saying so.
References
- Learning curve and surgical outcome of robotic assisted colorectal surgery with ERAS program
- Quantifying the impact of air pollution from coal-fired electricity generation on crop productivity in India
- En bloc resection of bladder tumour: the rebirth of past through reminiscence
- 3D Heads-Up Display vs. Standard Operating Microscope Vitrectomy for Rhegmatogenous Retinal Detachment
- Avoiding complications in endoscopic trans-sphenoidal surgery for pituitary adenoma: A beginner's perspective
- Bibliometric analysis of research trends and hotspots on robot-assisted thyroid surgery
- Delaying BCG immunotherapy onset after transurethral resection of non-muscle-invasive bladder cancer is associated with adverse survival outcomes
- Deep Survival: A Deep Cox Proportional Hazards Network
- Exploring the Limitations and Implications of the JIGSAWS Dataset for Robot-Assisted Surgery
- Preliminary investigation of the association between air pollution exposure and childhood asthma hospitalizations from 2015 to 2018 in East China