Clavien-Dindo classification

The Clavien-Dindo classification is an ordered severity classification for postoperative complications, based principally on the treatment needed to manage a deviation from the normal postoperative course. The 2004 proposal by Daniel Dindo, Nicolas Demartines and Pierre-Alain Clavien revised an earlier system and tested the revision using a 6,336-patient cohort and an international surgeon survey.^[ref]

The main grades summarize different management consequences. Grade I permits only the limited measures specified in the original definition. Grade II includes additional pharmacological treatment and the specified categories of transfusion or nutritional support. Grade III requires surgical, endoscopic or radiological intervention; IIIa and IIIb distinguish whether general anaesthesia is used. Grade IV concerns a life-threatening complication requiring intermediate or intensive care, with subdivisions for single- and multiple-organ dysfunction. Grade V denotes death. The original table supplies the precise exceptions and qualifications.^[ref]

The scale is ordered, not a ruler with equal intervals. Grade IV does not mean “twice as severe” as grade II. A numerical mean of grade labels consequently requires a justification beyond their ordering. Reporting the proportions in each category preserves information that a single average can hide.

Consider two illustrative groups of ten people. In each group, one person’s highest complication grade is III and the other nine have none. A highest-grade summary is identical. If the first affected person has one event and the second has three separate events, the summary still does not distinguish their total complication burden. Thus the unit of analysis, repeated-event rule and follow-up interval belong in the report. The original cohort analysis explicitly recorded the most severe complication when a patient had several.^[ref]

The classification makes outcome descriptions more reproducible, but assigning a grade is distinct from establishing why an event occurred or which care strategy would have prevented it. It is most informative alongside clear case definitions and the wider evaluation of perioperative care.

Surgical complication count versus patient incidence distinguishes the number of postoperative events from the number of people who experience at least one event. A person can have several postoperative complications. Counting events and counting affected patients therefore answer different questions, even with the same follow-up period.

Let $N$ be the number of operated patients with complete observation through a specified time, and let $K_i$ be the number of qualifying events for patient $i$. ThenHere $E$ is an event count, $\hat{p}$ estimates the proportion experiencing at least one complication during the specified period, and $E/N$ is mean events per patient. Unlike $\hat{p}$, the last quantity can exceed one.

For an original example, one 100-patient cohort has 40 events concentrated among 20 people. Another has 40 events, each in a different person. Both have $0.40$ events per patient, but their patient incidences are $20\%$ and $40\%$. Neither summary alone reveals severity, persistence or how closely events cluster in time.

The Clavien-Dindo classification adds severity information, but a highest-grade-per-patient summary still discards multiplicity. In the original validation paper, the most severe complication was registered when a patient had more than one. That is an explicit patient-level severity convention, rather than a count of every complication.^[ref]

Time and denominators must be specified. With 40 events over $3000$ patient-days, the event rate is $40/3000\approx 0.0133$ per patient-day; it is not a $1.33\%$ probability for each individual day. With incomplete follow-up, simple affected-patient proportions require further assumptions or methods accounting for observation time and competing events. Recurrent postoperative event analysis also has to account for dependence between several events in the same patient, rather than treating them as independent extra patients.

A reproducible comparison therefore defines the qualifying postoperative complication, observation window, patient denominator, event-counting rule and severity summary together. The original classification separately distinguishes complications from inherent sequelae and failure to cure, so those categories should not silently be added to the same numerator.^[ref]

Recurrent postoperative event analysis studies several qualifying events occurring in the same patient after surgery, rather than retaining only a first event or the highest severity grade. It requires an event definition that distinguishes a new episode from another observation of an ongoing episode. The original Clavien-Dindo classification validation deliberately registered the most severe complication for patients with several; a recurrent-event question requires more of the event history.^[ref]

Let $N_i(t)$ count qualifying events in patient $i$ by postoperative time $t$. This is a counting process. With complete observation of $n$ independent patients through $t$, the mean cumulative postoperative event count has the simple estimatorIt describes average accumulated burden. It is different from the probability $\mathrm{Pr}(N_i(t)\ge 1)$ of any event, and neither quantity specifies the severity of the events counted.

Within-patient dependence matters for uncertainty. At a fixed time, write $K_i=N_i(t)$ and let $s_K^2$ be the usual sample variance across patients. The patient-based standard error of the mean is $s_K/\sqrt{n}$ under independent-patient sampling. An original four-patient illustration with counts $(0,0,0,4)$ gives mean $1$, sample variance $4$ and estimated standard error $1$. Treating the total count as a Poisson count would instead give $\sqrt{4}/4=0.5$. This small example demonstrates different model assumptions; it does not validate either model or a normal confidence interval for four patients.

A proper record includes patient identity, episode times, resolution rules, severity and observation periods. Resampling patients with all their events intact can preserve within-patient dependence; resampling individual events as though they came from unrelated people cannot. An event intensity model may additionally condition on previous events, evolving health and exposure time, rather than assuming a constant rate.

Incomplete follow-up and death change the question. A rate among people currently alive and observed differs from total expected burden per original operated patient. Death prevents future postoperative events, whereas loss to follow-up hides whether events occurred. These require explicit observation and competing risks assumptions instead of an automatic independent-censoring treatment. The scientific aim is to measure repeated burden without confusing events, patients and time at risk.

The mean cumulative postoperative event count is the expected number of specified postoperative episodes accumulated per operated patient by a stated time. It measures repeated event burden, rather than only whether a patient had any event. An event definition must distinguish a new episode from continued observation of an existing complication.

Let $N_i(t)$ be patient $i$'s counting process for qualifying events up to postoperative time $t$. The population target isWith complete observation of $n$ independent patients through $t$, its empirical estimate isThis can exceed one. In contrast, the probability of at least one event is $p(t)=\mathrm{Pr}(N_i(t)\ge 1)$ and cannot exceed one. The inequality $\mu (t)\ge p(t)$ follows because each patient with any event contributes at least one to the count.

For an original five-patient example, counts $(0,0,0,0,5)$ give $\hat{\mu }=1$ and an observed any-event fraction $1/5$. Counts $(1,1,1,1,1)$ also give mean count $1$, but any-event fraction $1$. The same mean burden can describe very different distributions across patients. Neither count summary encodes the severity of each episode.

Uncertainty should respect the patient as the sampling unit. At fixed $t$, write $K_i=N_i(t)$ and calculateThe first example has $s_K^2=5$ and estimated standard error $1$. Treating its five events as five unrelated patient outcomes would discard the observed clustering. A small-sample estimated standard error is not itself a validated confidence interval.

Death and incomplete follow-up require explicit definitions. For accumulated burden per original operated patient, events cease at death but events already experienced remain counted; the denominator is not silently restricted to survivors. D. Ghosh and D. Y. Lin studied the marginal mean cumulative number of recurrent events in the presence of a terminal event, including nonparametric estimation and comparison.^[ref]

Loss to follow-up hides later events, whereas death prevents them. Simply averaging the available patients at each time can change the target population. Suitable observation-process assumptions and a censoring-aware method are needed for incomplete data. Comparisons should also show mortality: fewer accumulated events can accompany shorter survival, rather than reduced morbidity among otherwise comparable survivors.