Vrednotenje presejalnih programov za odkrivanje raka z analizo preživetja

Vratanar, Bor (author), Pohar Perme, Maja (mentor)

Presejalni programi za odkrivanje raka omogočajo, da z rednimi presejalnimi testi odkrijemo raka pri na videz zdravih posameznikih v zgodnejši fazi bolezni, ko je zdravljenje bolj učinkovito. Evalvacija presejalnih programov je pomembna za laično javnost, saj pomaga pri razumevanju prednosti in slabosti presejanja, in za odločevalce, ki morajo zagotoviti, da so sredstva v zdravstvu optimalno porabljena. V doktorski nalogi smo uspešnost presejalnih programov ovrednotili z analizo preživetja. Dosedanje raziskave na tem področju so poskušale oceniti učinkovitost zgodnjega zdravljenja tako, da so primerjale preživetje bolnikov, pri katerih smo odkrili raka na presejalnem testu (presejalno odkriti raki), s preživetjem bolnikov, pri katerih smo raka odkrili na podlagi simptomov, ki so se pojavili v intervalu med zadnjim in naslednjim presejalnim testom (intervalni raki). Gre za preprosto, a pristrano primerjavo, ki jo poimenujemo "naivna primerjava". Dosedanje raziskave so poskušale pristranost pri naivni primerjavi zmanjšati s pomočjo statističnih modelov, vendar so bile pri tem zgolj deloma uspešne. Trenutni pristopi, ki temeljijo na analizi preživetja, zato niso primerni za vrednotenje uspešnosti presejalnih programov. V doktorski nalogi smo v primerjavi z dosedanjo literaturo ubrali drugačen pristop. Da bi lahko natančno definirali primerjavo, ki nas zanima, in upoštevali vso kompleksnost problema, smo razvili novo, celovito notacijo. Ta temelji na notaciji potencialnih izidov in nam omogoča, da lahko potencialno izkušnjo iste osebe opišemo v dveh svetovih: v svetu 0 in svetu 1. V svetu 0 oseba ni povabljena v presejalni program, v svetu 1 pa je. Izid, ki nas zanima, je čas preživetja, definiran kot čas od diagnoze do smrti. Tako čas diagnoze kot čas smrti sta lahko različna v svetu 0 in svetu 1, kar pomeni, da ima oseba lahko različno preživetje glede na to, v katerem svetu jo opazujemo. Če želimo oceniti uspešnost presejalnih programov, moramo primerjati preživetje istih pacientov v svetu 1 (kjer so lahko zdravljeni predčasno), z njihovim preživetjem v svetu 0 (kjer so zdravljeni ob nastopu simptomov). Osredotočili smo se na podskupino presejalno odkritih pacientov, saj zgolj te paciente zdravimo predčasno v svetu 1 in se na ta način izognili pristranosti zaradi dolžine časa. Pomembno je tudi, da v obeh svetovih začnemo meriti čas do smrti ob istem datumu (imenovan čas nič), saj na ta način zagotovimo, da je primerjava resnično nepristrana. V našem primeru smo čas nič nastavili na datum diagnoze v svetu 0. V doktorski nalogi smo formalno definirali nepristrano primerjavo in pokazali, kako se razlikuje od naivne. Dokazali smo tudi, da so različni viri pristranosti, ki se pojavijo pri naivni primerjavi, med seboj aditivni. V praksi želimo uspešnost presejalnih programov oceniti na podlagi empiričnih podatkov. Preživetje presejalno odkritih pacientov, ki so zdravljeni predčasno, lahko ocenimo neposredno iz podatkov. Če želimo pri oceni preživetja prestaviti čas nič na datum, ko bi pacientom diagnosticirali raka na podlagi simptomov, lahko v ta namen uporabimo obstoječe parametrične metode, ki ocenjujejo pretečeni čas med diagnozo na podlagi presejalnega testa in diagnozo na podlagi simptomov (imenovan čas prednosti). Izziv, ki še ni bil rešen, pa je oceniti, kako bi presejalno odkriti pacienti živeli, če bi z zdravljenjem pričeli kasneje, ob simptomih - v praksi namreč vse paciente začnemo zdraviti takoj, ko postavimo diagnozo raka. V doktorski nalogi smo zato v ta namen razvili novo metodo. Predlagana metoda zahteva podatke za dve skupini oseb - za tiste, ki so vključeni v program in za tiste, ki niso. Najprej predpostavimo, da sta obe skupini randomizirani. Ideja metode je naslednja: bolnike, ki niso bili vabljeni v program, lahko razdelimo glede na tiste, ki bi v primeru vabila bili odkriti na presejanju in tiste, ki bi bili odkriti ob simptomih. Ne moremo sicer vedeti, kaj bi se zgodilo z vsakim posameznim bolnikom, lahko pa ocenimo delež vsake podskupine na podlagi podatkov skupine, ki je bila vabljena v presejalni program. Preživetje bolnikov, ki niso bili vabljeni v program tako lahko zapišemo kot uteženo vsoto obeh skupin. Preživetje združenih podskupin lahko neposredno opazujemo, hkrati pa predpostavimo tudi, da bi bilo preživetje simptomatsko odkritih rakov enako kot je pri tistih, ki so bili vabljeni programu, saj jim s programom ne moremo pomagati. Ker lahko ocenimo vse preostale dele utežene vsote, lahko torej izračunamo tisti del, ki nas zanima: preživetje bolnikov, ki bi lahko bili presejalno odkriti, v situaciji, ko jih pričnemo zdraviti šele ob simptomih. Našo metodo je mogoče združiti z obstoječimi metodami, kar nam omogoča, da ocenimo uspešnost zgodnjega zdravljenja brez pristranosti. V doktorskem delu smo izvedli različne simulacije, s katerimi smo želeli ponazoriti, kako različne nastavitve simulacije vplivajo na velikost pristranosti pri naivni primerjavi in s katerimi smo evalvirali predlagano metodo. Za ilustracijo naše metodologije smo evalvirali slovenski presejalni program za odkrivanje raka dojke (DORA) in pokazali, kako analizirati podatke v primeru, ko skupini vabljenih in nevabljenih nista primerljivi. Naši ilustrativni rezultati so podprli hipotezo, da zgodnje zdravljenje izboljša preživetje bolnic z rakom dojke. Vsebinski zaključki naše študije so omejeni, saj smo za oceno časa prednosti uporabili parametre iz predhodnih študij.

• Language: Slovenian
• Keywords: presejalni program, analiza preživetja, pristranost, potencialni izidi, rak dojke
• Work type: Doctoral dissertation
• Organization: MF - Faculty of Medicine
• Year: 2023
• PID: 20.500.12556/RUL-159039
• Publication date in RUL: 28.06.2024

Secondary language

• Language: English
• Title: Evaluating cancer screening programmes using survival analysis

Abstract

Cancer screening programmes (CSPs) routinely screen asymptomatic individuals in populations at risk of developing cancer using cost-effective tests. If a screening test yields positive results, a formal diagnostic test, such as a biopsy, is used to confirm the presence of cancer, ideally followed by early treatment. The underlying rationale is that treating patients who are detected on a screening test earlier than if they had presented with symptoms improves their probability of survival and, in some cases, prevents the development of an invasive malignancy altogether. Evaluation of CSPs is vital as it allows for balanced assessment of the benefits and harms associated with CSPs, and aids in optimizing public health resource allocation. Many different outcomes can be used for this purpose; in this thesis we focus on survival. To assess the effectiveness of early treatment, previous studies in this field relied on the naive comparison between interval (cancers detected based on symptoms after a negative test and prior to the next scheduled screening) and screen-detected cases (cancers detected based on a screening test). Realizing that the direct comparison between these two groups is subject to several biases, specifically lead time bias, length time bias and bias due to overdetection, several methods have been developed with the aim to reduce the bias. However, no existing approach has yet been able to fully account for these biases. Since the naive comparison is biased, we start this thesis by first defining the contrast of interest and subsequently search for suitable estimators. To this end, we develop a novel notation. The proposed notation is based on the notation of potential outcomes where we allow for each subject to be observed in two worlds: in world 1, we assume that the subject is invited to CSP and in world 0 we assume that the subject is not invited to CSP. Since the survival time is defined as the duration of time from diagnosis to death, both events are treated as potential outcomes and may therefore take different values in world 1 compared to world 0. In this thesis, the contrast of interest is defined as the comparison of survival for screen-detected cases in world 1, where cancer is detected earlier and treatment is initiated sooner (referred to as 'early treatment'), against their survival in world 0, where diagnosis occurs at symptom presentation and subsequent treatment is delayed compared to world 1 (referred to as 'delayed treatment'). To ensure a fair comparison, time zero is set at the same time in both worlds. In our case it was set at symptomatic detection, i.e. at cancer diagnosis in world 0. In this work, we formally define the contrast of interest and show how it is different from the naive comparison between interval and screen-detected cases. Lastly, we also provide the proof that the total bias that arises with the naive comparison is equal to the sum of lead time bias, length time bias and bias due to overdetection. With respect to the estimation, we show that the survival of screen-detected cases who received early treatment with time zero shifted to symptomatic detection, can be effectively estimated using existing parametric methods. However, we are missing an estimator that would allow us to estimate the second part of the contrast: the survival of screen-detected cases who received delayed treatment, as all screen-detected cases are invited for early treatment. To this end, we develop a new estimator that draws upon data from both cancer cases invited to the programme and those who were not; to simplify, we initially assume that the two groups are randomized. Our estimator is based on the idea that among cancer cases not invited to the programme, some would have been detected through screening if they had been invited, while others would in any case receive diagnoses based on symptoms, either due to non-attendance or as interval cases. While we cannot determine which patient among those not invited to the programme belongs to which subgroup, we can estimate the proportion of each hypothetical subgroup (screen-detected or detected based on symptoms) using data from cancer cases invited to the programme. This enables us to express the survival of cancer cases not invited to the programme as a weighted sum of the two hypothetical subgroups. The survival of screen-detected cases that would receive delayed treatment can therefore be expressed as a weighted difference between the overall survival of cancer cases not invited to the programme and the survival of cases not invited to the programme that would have been detected based on symptoms (assumed to be equal to the survival of symptomatic cases invited to the programme). By integrating the proposed estimator with existing methods, we show that it is possible to estimate the effectiveness of early treatment provided to screen-detected cases without any bias. Our work is supplemented by simulations and illustrated using empirical data. Simulations are performed to demonstrate how the size of each bias depends on the simulation parameters and to evaluate the proposed estimator. Data from Slovenian breast CSP are analysed to demonstrate our methodology and to show how the data can be matched if the two groups, invited vs. not invited, are not randomized. The illustrative results provided support for the hypothesis that early treatment improves probability of survival for screen-detected cases and revealed that length time bias is a major source of bias that should not be neglected. The conclusions should be interpreted in the light of the study’s limitations.

• Keywords: cancer screening, survival analysis, bias, counterfactual, breast cancer