Abstract
Background
Tobacco smoking after cancer diagnosis is associated with impaired survival but large real-world evidence across cancers accounting for treatment intent and key prognostic factors remains scarce. We evaluated associations between smoking and survival in a population-based cohort.
Materials and methods
We identified all patients who were diagnosed with solid cancers (International Classification of Diseases, 10th Revision codes C00-C43, C45-C68) and treated at two Finnish university hospitals (2009-2019). Smoking status (never, former, and current) was extracted from medical narratives using a validated natural language processing (NLP) pipeline linked to nationwide mortality data. Overall survival (OS) and cancer-specific survival (CSS) were estimated using Kaplan–Meier and multivariable Cox hazard ratios (HRs) adjusted for age, sex, performance status, comorbidity, nodal status, and treatment intent.
Results
Among 62 893 patients, smoking status was available for 45 395 (72%). Five-year OS was 71% in never, 58% in former, and 49% in current smokers; HR reference, 1.48 [95% confidence interval (CI) 1.43-1.53], and 1.91 (95% CI 1.85-1.97), respectively. Corresponding HRs for CSS were 1.52 (95% CI 1.47-1.59) and 2.09 (95% CI 2.02-2.17). Across most cancer types, 5-year CSS was 4%-19% lower in current versus never smokers. Impaired survival among current smokers (OS: HR 1.15-3.96; CSS: HR 1.12-4.07) and intermediate outcomes in former smokers were observed. Associations persisted after adjustment, including in the curative-intent subgroup (n = 36 936).
Conclusions
Current and former smoking are associated with impaired survival across multiple cancers, including after curative-intent treatment. These findings support systematic integration of smoking assessment and cessation into oncology care and demonstrate the utility of NLP for large-scale lifestyle outcome research.
Key words: artificial intelligence, big data, cancer, population-based, prognosis, smoking
Highlights
-
•
Smoking status was identified in 45 395 of 62 893 patients (72%) in a population-based cohort of solid cancers.
-
•
Current versus never smokers had poorer OS (HR 1.15-3.98) and CSS (HR 1.15-4.07) across most cancers.
-
•
Former smoking was also linked to poorer OS (HR 1.15-2.64 compared with never smokers) in many cancer types.
-
•
Impaired survival among smokers persisted after adjustment for major prognostic factors and curative-intent treatment.
-
•
NLP of routine health records enables large-scale research on smoking and cancer outcomes.
Introduction
As advances in cancer detection and treatment have substantially improved survival over recent decades, variation in long-term outcomes increasingly reflects factors beyond oncological therapy itself, including modifiable lifestyle behaviours such as cigarette smoking.1, 2, 3, 4 Large epidemiological studies and systematic reviews have demonstrated impaired overall survival (OS) and cancer-specific survival (CSS) among smokers compared with never smokers in several cancer types, including breast, prostate, colorectal, lung, bladder, and head and neck squamous-cell cancer (HNSCC), and melanoma.5, 6, 7, 8, 9, 10, 11, 12 However, results remain inconsistent for kidney and uterine cancers.13, 14, 15
Smoking also contributes to increased cardiovascular mortality, second primary cancers, and treatment-related complications among cancer survivors.16, 17, 18 Smoking during surgery, radiotherapy, and systemic anticancer therapy (SACT) impairs treatment efficacy and increases treatment-related toxicity.1,2,10 Several mechanisms have been proposed to explain these effects, including smoking-induced tissue hypoxia and impaired wound healing affecting surgical and radiotherapy outcomes, nicotine-mediated activation of cell-survival signalling pathways that inhibit apoptosis, and altered drug metabolism that may influence response to SACT.1, 2, 3,10,19
Despite substantial evidence linking smoking to poorer cancer outcomes, important research gaps remain. Most studies have focused on individual cancer types or selected treatment populations, and relatively few have simultaneously accounted for treatment intent, performance status, comorbidity burden, and other key prognostic factors when evaluating survival.2,5, 6, 7, 8, 9, 10, 11, 12, 13, 14 In addition, smoking status is frequently incompletely recorded in clinical/trial datasets and often relies on questionnaires with variable completion rates, while large population-based cohorts with systematically extracted smoking data combined with detailed clinical variables remain scarce.2, 3, 4, 5,12,14,20
This study aimed to quantify the impact of smoking status (never, former, and current) on OS and CSS after cancer diagnosis in a large population-based cohort of patients with solid cancers, using smoking information extracted from electronic health records with a natural language processing (NLP) pipeline, and to determine whether the association persists after accounting for treatment intent and major clinical prognostic factors.
Materials and methods
Health care organisation
Turku and Tampere University Hospitals are accredited cancer centres serving primary catchment populations of ∼480 000 and 540 000 inhabitants, respectively, together covering about 18% of the Finnish population. In the Finnish publicly funded health care system, oncological care is organised for geographically defined populations, and Turku and Tampere University Hospitals provide cancer surgery, radiotherapy, and SACT for their own catchment areas.
Radiotherapy and intravenous SACT given to patients in these catchment areas are solely delivered at the university hospitals and are therefore captured in the hospital records. Oral SACT is prescribed through hospital oncology units and likewise recorded. However, hormonal therapies frequently used in breast and prostate cancers are mainly dispensed through outpatient pharmacies and cannot be captured.
During the study period, some complex oncological surgical procedures were nationally centralised to university hospitals, including those for HNSCC, gastro-oesophageal, rectal, uterine, and pancreatic cancers, thus enriching the cohort with patients from outside the primary populations. In addition, the majority of palliative radiotherapies were organised at university hospitals, affecting proportion of palliative patients.
Cohort identification
All patients diagnosed with solid cancers [International Classification of Diseases, 10th Revision (ICD-10) codes C00-C68, excluding C44 due to low expected mortality] between 2009 and 2019 were identified. Lung cancer was classified as non-small-cell (NSCLC, C34.x1-5), small-cell (SCLC, C34.x6-7), or undefined (C34.x0/9). The index cancer was defined as the first treated cancer documented after 1 January 2009 in Turku and after 1 January 2010 in Tampere (Figure 1). Duplicate individuals were identified using their unique Finnish personal identity code; the first treated cancer was selected as the index cancer and treatment records from both hospitals were combined.
Figure 1.
Data sources and study population of the current study. ICD-10, International Classification of Diseases, 10th Revision.
Dates and underlying causes of death were obtained from Statistics Finland, which maintains the nationwide cause-of-death registry with validated high-quality death certificate data.21,22
Data validation
Manual diagnostic validation was carried out by comparing ICD-10 codes, causes of death, Nordic Classification of Surgical Procedures codes, SACT information, and radiotherapy data for each patient. The most common discrepancies involved lymphomas initially diagnosed as solid cancers and lung metastases diagnosed as primary tumours. For patients presenting simultaneously with multiple cancers within 3 months or with conflicting diagnostic information, the actual treatment administered was used to define the index cancer.
Smoking classification
In both hospitals, smoking status was extracted from free-text entries in the electronic medical records from 2009 to 2019 using an NLP pipeline combining rule-based extraction and machine-learning classification. On average, seven independent entries (range 1-116) describing smoking status were extracted per patient. Tobacco smoking-related phrases and sentences were first identified using regular expressions, after which patients were classified as never, former, or current smokers using classification models previously validated in Turku.20 Manual validation in Tampere showed a phrase-level precision of 0.93 (Supplementary Material, available at https://doi.org/10.1016/j.esmorw.2026.100791). At Turku, 92% of smoking phrases were during year of cancer diagnosis or after, compared with 58% at Tampere. No specific time point was used; thus, the NLP classifies patients based on the entire observational period. Former smokers therefore included all patients who had quit smoking during 2009-2019 or at any time before.
Baseline variables
Structured baseline variables recorded closest to the index cancer included age, sex, Eastern Cooperative Oncology Group performance status (ECOG PS), Charlson comorbidity index (CCI) (Supplementary Methods, available at https://doi.org/10.1016/j.esmorw.2026.100791), alcohol abuse (ICD code F10), and body mass index (BMI). Cancer stage (I-IV) or tumour–node–metastasis (TNM, version used during the diagnosis) classification of the index cancer was extracted directly from the electronic medical records, collecting the closest ‘stage X’ and ‘TxNxMx’ phrase to the index cancer date, which was available for 63% of patients, and reclassified as lymph node-negative, positive, or metastatic disease.
Treatment intent classification
Treatment intent was classified hierarchically into ‘curative-intent’, ‘palliative-intent’, or ‘undefined’ subgroups based on highest intensity of treatment given (Figure 1). Resections of the primary tumour, with or without definitive, post-operative, or (chemo)radiotherapy, and selected adjuvant chemotherapies were considered curative intent, excluding patients with stage IV disease. Any palliative radiotherapy or SACT was classified as palliative intent. The patients receiving best supportive care (BSC) only were classified in the palliative-intent subgroup. The remaining patients were assigned to the undefined group, which most likely included those treated only with hormonal therapy or clinical follow-up.
Statistics
Baseline variables were compared using chi-square with Bonferroni adjustment for multiple comparisons as appropriate. Two-sided P values <0.05 were considered statistically significant. For multivariable analysis, missing data were coded as ‘not available’ and the absence of a comorbidity code was interpreted as the absence of comorbidity. OS was defined as the time from cancer diagnosis (ICD code) to death from any cause, with patients alive at the end of follow-up censored on 31 December 2022. CSS was defined from cancer diagnosis to death with cancer recorded as the underlying cause of death,21,22 with patients alive on 31 December 2022 or dying from noncancer causes censored at the corresponding date. The reverse Kaplan–Meier method was used to calculate median follow-up. Survival was estimated using the Kaplan–Meier method and hazard ratios (HRs) with 95% confidence intervals (CIs) were calculated using Cox proportional hazards models. Adjustment for age, sex, ECOG PS, CCI, alcohol abuse, treatment intent, and nodal status was carried out with Cox multivariable regression, and potential selection bias due to missing smoking information was assessed using stabilised inverse probability of observation weighting (IPW). IBM SPSS Statistics version 26 was used (IBM, Armonk, NY) for all analyses, except the IPW analysis, which was carried out using R version 4.4.1 (R Foundation, Vienna, Austria).
Results
Patient characteristics and data completeness
A total of 62 893 patients with solid cancers were identified, of whom 30 999 were female and 31 894 male. Smoking status was available for 45 395 patients (72% of the cohort). The highest prevalence (37%-47%) of missing smoking status was observed among patients with uterine cancer or melanoma, while in other cancer groups, missing smoking status ranged from 8% to 35% (Table 1).
Table 1.
Association between smoking status and postdiagnosis overall and CSS
| Smoking status by cancer group | n (%) | Median OS, months (5-year rate) | HR for OS (95% CI) | HR for CSS (95% CI) | Adjusteda HR for CSS (95% CI) |
|---|---|---|---|---|---|
| All cancers (C00-C43, C45-C68, N = 62 893) | |||||
| Never | 22 742 (36%) | 151 (71%) | Reference | Reference | Reference |
| Former | 10 852 (17%) | 98 (58%) | 1.48 (1.43-1.53) | 1.52 (1.47-1.59) | 1.32 (1.26-1.37) |
| Current | 11 801 (19%) | 60 (49%) | 1.91 (1.85-1.97) | 2.09 (2.02-2.17) | 1.80 (1.73-1.87) |
| Not available | 17 498 (28%) | 105 (60%) | 1.44 (1.39-1.48) | 1.43 (1.39-1.49) | 1.35 (1.30-1.40) |
| Head and neck (C00-C14, C30-C32, n = 2817) | |||||
| Never | 734 (26%) | 133 (67%) | Reference | Reference | Reference |
| Former | 703 (25%) | 116 (65%) | 1.03 (0.88-1.20) | 1.06 (0.88-1.28) | 1.05 (0.87-1.27) |
| Current | 966 (34%) | 53 (48%) | 1.81 (1.58-2.07) | 1.81 (1.54-2.13) | 1.73 (1.45-2.05) |
| Not available | 414 (15%) | 60 (50%) | 1.61 (1.37-1.90) | 1.39 (1.13-1.71) | 1.34 (1.08-1.65) |
| Gastro-oesophageal (C15-C16, n = 1973) | |||||
| Never | 552 (28%) | 14 (25%) | Reference | Reference | Reference |
| Former | 414 (21%) | 12 (20%) | 1.10 (0.96-1.27) | 1.10 (0.95-1.28) | 1.13 (0.97-1.32) |
| Current | 504 (26%) | 9 (16%) | 1.35 (1.18-1.54) | 1.36 (1.18-1.56) | 1.42 (1.22-1.65) |
| Not available | 503 (26%) | 8 (21%) | 1.24 (1.08-1.41) | 1.22 (1.06-1.41) | 1.11 (0.95-1.28) |
| Colorectal (C18-C20, n = 6447) | |||||
| Never | 2546 (39%) | 105 (62%) | Reference | Reference | Reference |
| Former | 1127 (17%) | 82 (57%) | 1.15 (1.05-1.27) | 1.12 (1.00-1.25) | 1.05 (0.93-1.18) |
| Current | 929 (14%) | 75 (54%) | 1.25 (1.13-1.38) | 1.22 (1.08-1.37) | 1.18 (1.05-1.34) |
| Not available | 1845 (29%) | 64 (51%) | 1.45 (1.34-1.57) | 1.37 (1.24-1.51) | 1.23 (1.11-1.36) |
| Pancreatic (C25, n = 2129) | |||||
| Never | 682 (32%) | 6 (10%) | Reference | Reference | Reference |
| Former | 323 (15%) | 8 (9%) | 0.96 (0.84-1.10) | 0.95 (0.82-1.09) | 0.95 (0.82-1.09) |
| Current | 485 (23%) | 5 (6%) | 1.15 (1.02-1.30) | 1.15 (1.02-1.30) | 1.14 (1.01-1.30) |
| Not available | 639 (30%) | 4 (6%) | 1.28 (1.15-1.43) | 1.25 (1.12-1.40) | 1.09 (0.97-1.23) |
| NSCLC (C34, n = 3710) | |||||
| Never | 466 (13%) | 18 (27%) | Reference | Reference | Reference |
| Former | 1155 (31%) | 12 (20%) | 1.22 (1.09-1.38) | 1.24 (1.09-1.40) | 1.13 (0.99-1.29) |
| Current | 1746 (47%) | 7 (12%) | 1.59 (1.42-1.78) | 1.62 (1.43-1.83) | 1.44 (1.27-1.64) |
| Not available | 343 (9%) | 5 (11%) | 1.74 (1.50-2.03) | 1.76 (1.50-2.06) | 1.23 (1.07-1.49) |
| SCLC (C34, n = 727) | |||||
| Never | 66 (9%) | 115 (57%) | Reference | Reference | Reference |
| Former | 170 (23%) | 12 (21%) | 2.64 (1.79-3.89) | 2.76 (1.83-4.14) | 1.67 (1.10-2.55) |
| Current | 436 (60%) | 8 (9%) | 3.98 (2.75-5.75) | 4.07 (2.76-5.99) | 2.14 (1.43-3.21) |
| Not available | 55 (8%) | 8 (16%) | 3.44 (2.19-5.39) | 3.46 (2.16-5.54) | 1.66 (1.01-2.72) |
| Cutaneous melanoma (C43, n = 3234) | |||||
| Never | 1290 (40%) | 168 (79%) | Reference | Reference | Reference |
| Former | 379 (12%) | 163 (70%) | 1.30 (1.08-1.57) | 1.33 (1.05-1.68) | 1.03 (0.81-1.31) |
| Current | 362 (11%) | 150 (72%) | 1.30 (1.08-1.58) | 1.44 (1.14-1.83) | 1.46 (1.15-1.86) |
| Not available | 1203 (37%) | Not reached (73%) | 1.16 (1.01-1.33) | 1.03 (0.86-1.23) | 1.37 (1.14-1.64) |
| Breast (C50, n = 13 120) | |||||
| Never | 7281 (55%) | 171 (87%) | Reference | Reference | Reference |
| Former | 1741 (13%) | 174 (89%) | 0.81 (0.72-0.92) | 0.85 (0.72-0.99) | 0.96 (0.82-1.12) |
| Current | 1792 (14%) | 171 (85%) | 1.06 (0.96-1.18) | 1.10 (0.95-1.26) | 1.28 (1.11-1.48) |
| Not available | 2306 (18%) | 145 (70%) | 2.16 (1.99-2.33) | 2.41 (2.17-2.67) | 1.76 (1.58-1.96) |
| Uterine (C54-C55, n = 2749) | |||||
| Never | 1103 (40%) | 157 (76%) | Reference | Reference | Reference |
| Former | 134 (5%) | 168 (72%) | 1.05 (0.78-1.42) | 0.92 (0.63-1.34) | 1.11 (0.75-1.63) |
| Current | 222 (8%) | Not reached (79%) | 0.77 (0.59-1.01) | 0.76 (0.55-1.05) | 0.91 (0.65-1.26) |
| Not available | 1290 (47%) | 156 (72%) | 1.16 (1.01-1.32) | 1.07 (0.91-1.25) | 1.08 (0.91-1.27) |
| Prostate (C61, n = 13 207) | |||||
| Never | 4072 (31%) | 158 (83%) | Reference | Reference | Reference |
| Former | 2722 (21%) | 139 (81%) | 1.22 (1.12-1.32) | 1.23 (1.09-1.39) | 1.08 (0.95-1.22) |
| Current | 1825 (14%) | 134 (75%) | 1.42 (1.30-1.55) | 1.62 (1.43-1.85) | 1.69 (1.48-1.93) |
| Not available | 4588 (35%) | 171 (80%) | 0.99 (0.92-1.07) | 1.06 (0.95-1.18) | 1.43 (1.28-1.61) |
| Kidney (C64, n = 2023) | |||||
| Never | 666 (33%) | 101 (61%) | Reference | Reference | Reference |
| Former | 304 (15%) | 64 (52%) | 1.22 (1.02-1.47) | 1.22 (0.99-1.50) | 1.10 (0.88-1.37) |
| Current | 352 (17%) | 80 (58%) | 1.15 (0.96-1.37) | 1.10 (0.89-1.36) | 1.04 (0.84-1.29) |
| Not available | 701 (35%) | 71 (53%) | 1.31 (1.13-1.51) | 1.29 (1.09-1.53) | 1.49 (1.25-1.78) |
| Urinary tract (C65-C68, n = 2676) | |||||
| Never | 635 (24%) | 124 (66%) | Reference | Reference | Reference |
| Former | 604 (23%) | 72 (57%) | 1.48 (1.27-1.73) | 1.48 (1.20-1.81) | 1.40 (1.13-1.73) |
| Current | 638 (24%) | 85 (61%) | 1.30 (1.11-1.52) | 1.36 (1.11-1.67) | 1.66 (1.34-2.06) |
| Not available | 799 (30%) | 69 (53%) | 1.55 (1.34-1.79) | 1.63 (1.34-1.97) | 1.77 (1.46-2.15) |
Statistically significant results in bold.
CCI, Charlson comorbidity index; CI, confidence interval; CSS, cancer-specific survival; ECOG PS, Eastern Cooperative Oncology Group performance status; HR, hazard ratio; NSCLC, non-small-cell lung cancer; OS, overall survival; SCLC, small-cell lung cancer.
Multivariable Cox regression adjusted for age >75 years, sex (where applicable), ECOG PS (0-1 versus 2-4 versus not available), CCI (0 versus 1 versus 2-10), alcohol abuse diagnosis, and treatment intent (curative intent versus palliative intent versus undefined).
Patients with unknown smoking status were typically older, had poorer ECOG PS, and had fewer recorded comorbidities (likely reflecting underreporting rather than a true absence of comorbidity), while sex and BMI distribution were similar. Missing staging and treatment intent were more frequently observed among patients with missing smoking status (Supplementary Table S1, available at https://doi.org/10.1016/j.esmorw.2026.100791). Missing smoking status was more common in Tampere but decreased over the study period from 37% to 28%, along with other missing data (Supplementary Table S2, available at https://doi.org/10.1016/j.esmorw.2026.100791).
Annual cancer incidence increased over the study period (Supplementary Table S3, available at https://doi.org/10.1016/j.esmorw.2026.100791) and was comparable with national online statistics from the Finnish Cancer Registry, indicating high completeness of case capture. Centralisation of HNSCC and uterine cancer treatment resulted in higher-than-expected capture rates, while capture was otherwise largely complete, except for melanoma, for which primary surgery is also carried out outside university hospitals, and urinary tract cancers in Tampere. Consequently, a slightly higher-than-expected proportion of metastatic disease23 and poorer survival compared with nationwide data were observed (Supplementary Table S3, available at https://doi.org/10.1016/j.esmorw.2026.100791).
Association between smoking status and survival
After a median follow-up of 99 months (95% CI 98-100) and a minimum follow-up of 36 months, 29 943 deaths had occurred. Among all patients with solid cancers, both current and former smokers had impaired OS, with HRs of 1.91 (95% CI 1.85-1.97) and 1.48 (95% CI 1.43-1.53), respectively, compared with never smokers (Figure 2A). Current smokers had impaired OS across most cancer types, with HRs ranging from 1.15 to 3.98, with the exception of breast, uterine, and kidney cancers (Table 1); delta for median OS varied between 1 month (pancreatic cancer) and 107 months (SCLC), and absolute differences for 5-year OS rates were between 4% and 19%. Impaired CSS (Figure 2B) was consistently observed among current smokers compared with never smokers (HR 1.15-4.07; Table 1) except for breast, uterine, and kidney cancers.
Figure 2.
Survival among patients with solid cancer (C00-C68, excluding C44). OS (A) and CSS (B) in all patients with known smoking status, and among patients with curative-intent treatment only (C) and (D), respectively. CI, confidence interval; CSS, cancer-specific survival; HR, hazard ratio; OS, overall survival.
After adjustment for age, sex, treatment intent, alcohol abuse, comorbidity index, and ECOG PS, impaired CSS was observed for all solid cancers including breast cancer but still not for uterine and kidney cancers (adjusted HR 1.14-2.14; Table 1 and Supplementary Table S4, available at https://doi.org/10.1016/j.esmorw.2026.100791).
Cessation year was documented for 6279 of 10 852 former smokers, of whom 1131 (18%) quit around the time of cancer diagnosis (Supplementary Table S5, available at https://doi.org/10.1016/j.esmorw.2026.100791). Among former smokers compared with never smokers, impaired OS (HR range 1.15-2.64) and CSS (unadjusted HR 1.12-2.76 and adjusted HR 1.32-1.67) were observed in all patients and in subgroups of colorectal, lung, melanoma, prostate, and urinary tract cancers, while the other cancer groups were more comparable to never smokers (HR 0.81-1.22 for OS or CSS; Table 1). Patients with unknown smoking status had shorter survival than those with known smoking status, median OS 105 compared with 116 months with HR 1.03 (95% CI 1.02-1.04), although differences varied by cancer type (Table 1).
Compared with never smokers, at baseline, current smokers were uniformly 3-6 years younger, more often male, less frequently allocated to curative-intent treatment, and had a higher burden of comorbidities, metastatic disease, more frequent alcohol abuse, low BMI, and poorer ECOG PS (Table 2).
Table 2.
Baseline characteristics by smoking status within each cancer group
| Smoking status by cancer group | Mean age (IQR), years | Sex: male versus female | ECOG PS: 2-4 versus 0-1 | CCI: 1 or more versus 0 | Alcohol abuse: yes versus no | BMI <20 versus ≥20 kg/m2 | Curative versus other intent treatment | Metastatic (stage IV) versus localised |
|---|---|---|---|---|---|---|---|---|
| All cancers (C00-C43, C45-C68) | ||||||||
| Never, n = 22 742 | 69 (61-77) | 8462 (37%) | 2438 (16%) | 8435 (37%) | 287 (1%) | 782 (5%) | 15 313 (67%) | 2478 (16%) |
| Former, n = 10 852 | 69 (62-76) | 7438 (69%) | 1548 (20%) | 5383 (50%) | 304 (3%) | 306 (4%) | 6533 (60%) | 1533 (22%) |
| Current, n = 11 801 | 65 (58-72) | 7117 (60%) | 1976 (26%) | 5664 (48%) | 940 (8%) | 941 (11%) | 6135 (52%) | 2231 (29%) |
| P value | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 |
| Head and neck (C00-C14, C30-C32) | ||||||||
| Never, n = 735 | 68 (60-78) | 339 (46%) | 99 (21%) | 239 (32%) | 8 (1%) | 31 (5%) | 650 (88%) | 62 (9%) |
| Former, n = 703 | 66 (58-74) | 547 (78%) | 93 (18%) | 286 (41%) | 32 (5%) | 26 (5%) | 626 (89%) | 60 (9%) |
| Current, n = 966 | 63 (57-69) | 724 (75%) | 222 (33%) | 383 (40%) | 128 (13%) | 137 (19%) | 814 (84%) | 80 (9%) |
| P value | <0.001 | <0.001 | <0.001 | 0.003 | <0.001 | <0.001 | 0.007 | 0.73 |
| Gastro-oesophageal (C15-C16) | ||||||||
| Never, n = 552 | 71 (63-75) | 277 (50%) | 133 (31%) | 265 (48%) | 16 (3%) | 59 (13%) | 185 (34%) | 135 (47%) |
| Former, n = 414 | 69 (62-74) | 333 (80%) | 101 (30%) | 203 (49%) | 12 (3%) | 24 (7%) | 168 (41%) | 107 (47%) |
| Current, n = 504 | 65 (58-70) | 399 (79%) | 140 (38%) | 244 (48%) | 56 (11%) | 100 (26%) | 168 (33%) | 136 (54%) |
| P value | <0.001 | <0.001 | 0.05 | 0.16 | <0.001 | <0.001 | 0.007 | 0.19 |
| Colorectal (C18-C20) | ||||||||
| Never, n = 2546 | 70 (64-78) | 1110 (44%) | 299 (18%) | 1124 (44%) | 31 (1%) | 129 (6%) | 1936 (76%) | 483 (21%) |
| Former, n = 1127 | 70 (64-76) | 855 (76%) | 162 (20%) | 602 (53%) | 28 (3%) | 39 (4%) | 831 (74%) | 242 (24%) |
| Current, n = 929 | 67 (61-73) | 580 (62%) | 130 (22%) | 488 (52%) | 68 (7%) | 89 (12%) | 643 (69%) | 213 (27%) |
| P value | <0.001 | <0.001 | 0.07 | <0.001 | <0.001 | <0.001 | <0.001 | 0.004 |
| Pancreatic (C25) | ||||||||
| Never, n = 682 | 72 (63-75) | 264 (39%) | 177 (34%) | 389 (57%) | 6 (1%) | 43 (8%) | 142 (21%) | 224 (69%) |
| Former, n = 323 | 69 (59-71) | 222 (69%) | 85 (34%) | 192 (59%) | 8 (3%) | 13 (5%) | 75 (23%) | 127 (73%) |
| Current, n = 485 | 66 (59-71) | 309 (64%) | 127 (37%) | 287 (59%) | 38 (8%) | 50 (15%) | 76 (16%) | 194 (76%) |
| P value | <0.001 | <0.001 | 0.63 | 0.24 | <0.001 | <0.001 | 0.06 | 0.14 |
| NSCLC (C34) | ||||||||
| Never, n = 466 | 74 (64-76) | 131 (28%) | 95 (28%) | 194 (42%) | 3 (1%) | 19 (6%) | 112 (24%) | 183 (57%) |
| Former, n = 1155 | 72 (64-76) | 837 (73%) | 291 (33%) | 674 (58%) | 29 (3%) | 64 (7%) | 266 (23%) | 365 (47%) |
| Current, n = 1747 | 69 (61-71) | 1188 (68%) | 487 (39%) | 1044 (60%) | 136 (8%) | 169 (14%) | 268 (15%) | 711 (59%) |
| P value | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 |
| SCLC (C34) | ||||||||
| Never, n = 66 | 69 (62-80) | 17 (26%) | 9 (23%) | 24 (36%) | 0 | 0 | 27 (41%) | 21 (31%) |
| Former, n = 170 | 69 (63-75) | 110 (65%) | 46 (32%) | 81 (48%) | 5 (3%) | 9 (6%) | 22 (13%) | 65 (66%) |
| Current, n = 436 | 66 (61-71) | 255 (59%) | 138 (38%) | 239 (55%) | 35 (8%) | 56 (15%) | 22 (5%) | 97 (78%) |
| P value | 0.002 | <0.001 | 0.08 | 0.06 | 0.006 | 0.004 | <0.001 | <0.001 |
| Cutaneous melanoma (C43) | ||||||||
| Never, n = 1290 | 65 (58-76) | 587 (46%) | 104 (16%) | 488 (38%) | 18 (1%) | 33 (4%) | 824 (64%)a | 75 (8%) |
| Former, n = 379 | 67 (61-75) | 292 (77%) | 44 (20%) | 193 (51%) | 13 (3%) | 5 (2%) | 239 (63%)a | 36 (13%) |
| Current, n = 362 | 60 (51-71) | 223 (62%) | 38 (19%) | 135 (47%) | 20 (6%) | 8 (3%) | 215 (59%)a | 43 (18%) |
| P value | <0.001 | <0.001 | 0.31 | <0.001 | <0.001 | 0.05 | 0.13 | <0.001 |
| Breast (C50) | ||||||||
| Never, n = 7281 | 65 (55-73) | 31 (<1%) | 741 (11%) | 2020 (28%) | 47 (1%) | 231 (4%) | 6079 (84%) | 380 (7%) |
| Former, n = 1741 | 62 (55-69) | 15 (1%) | 155 (10%) | 499 (29%) | 27 (2%) | 45 (3%) | 1522 (87%) | 71 (5%) |
| Current, n = 1792 | 60 (53-67) | 8 (<1%) | 163 (10%) | 543 (30%) | 68 (5%) | 114 (9%) | 1501 (84%) | 90 (7%) |
| P value | <0.001 | 0.06 | 0.06 | 0.08 | <0.001 | <0.001 | <0.001 | 0.06 |
| Uterine (C54-C55) | ||||||||
| Never, n = 1103 | 69 (63-77) | n/a | 74 (15%) | 365 (33%) | 4 (1%) | 30 (4%) | 969 (88%) | 72 (7%) |
| Former, n = 134 | 67 (62-72) | n/a | 12 (20%) | 62 (46%) | 0 | 3 (3%) | 121 (90%) | 8 (7%) |
| Current, n = 222 | 65 (58-70) | n/a | 7 (9%) | 84 (37%) | 4 (2%) | 6 (3%) | 199 (90%) | 15 (7%) |
| P value | <0.001 | 0.23 | 0.02 | 0.04 | 0.64 | 0.49 | 0.96 | |
| Prostate (C61) | ||||||||
| Never, n = 4072 | 71 (64-74) | n/a | 219 (11%) | 1750 (43%) | 104 (3%) | 50 (2%) | 2555 (63%) | 184 (8%) |
| Former, n = 2722 | 71 (64-74) | n/a | 214 (13%) | 1456 (53%) | 86 (3%) | 29 (2%) | 1725 (63%) | 123 (8%) |
| Current, n = 1825 | 68 (61-71) | n/a | 142 (17%) | 951 (52%) | 179 (10%) | 49 (4%) | 1077 (59%) | 110 (11%) |
| P value | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | 0.04 | |
| Kidney (C64) | ||||||||
| Never, n = 666 | 69 (60-78) | 330 (50%) | 91 (29%) | 350 (52%) | 11 (2%) | 18 (3%) | 453 (68%) | 117 (32%) |
| Former, n = 304 | 69 (62-74) | 250 (82%) | 44 (25%) | 185 (61%) | 10 (3%) | 9 (3%) | 188 (62%) | 74 (45%) |
| Current, n = 352 | 63 (55-71) | 247 (70%) | 52 (34%) | 204 (58%) | 32 (9%) | 17 (6%) | 221 (63%) | 72 (36%) |
| P value | <0.001 | <0.001 | 0.19 | 0.03 | <0.001 | 0.05 | 0.17 | 0.02 |
| Urinary tract (C66-C68) | ||||||||
| Never, n = 635 | 73 (66-82) | 428 (67%) | 58 (28%) | 317 (50%) | 9 (1%) | 23 (5%) | 515 (81%) | 23 (16%) |
| Former, n = 604 | 74 (68-80) | 546 (90%) | 81 (32%) | 396 (66%) | 18 (3%) | 9 (2%) | 514 (85%) | 33 (18%) |
| Current, n = 638 | 68 (62-76) | 531 (83%) | 65 (33%) | 390 (61%) | 33 (5%) | 35 (7%) | 529 (83%) | 31 (18%) |
| P value | <0.001 | <0.001 | 0.50 | <0.001 | <0.001 | <0.001 | 0.07 | 0.83 |
Patients with missing data are excluded.
BMI, body mass index; CCI, Charlson comorbidity index; CSS, cancer-specific survival; IQR, interquartile range; ECOG PS, Eastern Cooperative Oncology Group performance status; n/a, not applicable; NSCLC, non-small-cell lung cancer; SCLC, small-cell lung cancer.
Represents melanoma resections with sentinel node biopsy.
IPW-derived HRs were similar to the primary analyses, indicating that selective availability of smoking information did not markedly bias the main findings. Detailed weighted analyses, including truncation at the 1st and 99th percentiles, are shown in Supplementary Table S6 (available at https://doi.org/10.1016/j.esmorw.2026.100791).
Survival in patients treated with curative intent
In general, the impaired OS and CSS effect of smoking was observed among the subgroup of 36 936 patients treated with curative intent (Figure 2C and D, Table 3), except for breast, uterine, and kidney cancers. After adjustment for nodal status, CSS was also impaired in breast cancer but no longer in melanoma.
Table 3.
Survival among patients allocated to curative-intent treatment with known smoking status
| Smoking status by cancer group | n | Median OS in months (5-year rate) | HR for OS (95% CI) | HR for CSS (95% CI) | Adjusteda HR for CSS (95% CI) |
|---|---|---|---|---|---|
| All cancers (C00-C43, C45-C68, N = 27 981) | |||||
| Never | 15 313 | 170 (82%) | Reference | Reference | Reference |
| Former | 6533 | 161 (77%) | 1.30 (1.23-1.37) | 1.30 (1.22-1.39) | 1.17 (1.08-1.27) |
| Current | 6135 | 149 (71%) | 1.57 (1.49-1.65) | 1.62 (1.52-1.72) | 1.67 (1.55-1.81) |
| Head and neck (C00-C14, C30-C32, n = 2090) | |||||
| Never | 650 | 144 (70%) | Reference | Reference | Reference |
| Former | 626 | 133 (70%) | 1.01 (0.85-1.20) | 1.05 (0.85-1.29) | 1.08 (0.87-1.35) |
| Current | 814 | 62 (51%) | 1.87 (1.61-2.17) | 1.85 (1.54-2.22) | 1.91 (1.57-2.33) |
| Gastro-oesophageal (C15-C16, n = 521) | |||||
| Never | 185 | 67 (52%) | Reference | Reference | Reference |
| Former | 168 | 32 (38%) | 1.47 (1.16-2.00) | 1.57 (1.17-2.12) | 1.48 (1.08-1.99) |
| Current | 168 | 33 (36%) | 1.56 (1.20-2.05) | 1.64 (1.21-2.20) | 1.76 (1.29-2.40) |
| Colorectal (C18-C20, n = 3410) | |||||
| Never | 1936 | 130 (71%) | Reference | Reference | Reference |
| Former | 831 | 115 (69%) | 1.11 (0.98-1.26) | 1.05 (0.90-1.24) | 0.98 (0.83-1.15) |
| Current | 643 | 102 (63%) | 1.24 (1.09-1.42) | 1.21 (1.03-1.45) | 1.24 (1.04-1.47) |
| Pancreatic (C25, n = 293) | |||||
| Never | 147 | 37 (34%) | Reference | Reference | Reference |
| Former | 78 | 28 (24%) | 1.15 (0.84-1.57) | 1.09 (0.78-1.53) | 1.15 (0.80-1.65) |
| Current | 71 | 24 (21%) | 1.70 (1.24-2.33) | 1.38 (1.00-1.92) | 1.45 (1.15-2.03) |
| NSCLC (C34, n = 646) | |||||
| Never | 112 | 99 (68%) | Reference | Reference | Reference |
| Former | 266 | 79 (58%) | 1.37 (0.99-1.90) | 1.48 (1.02-2.16) | 1.22 (0.82-1.81) |
| Current | 268 | 57 (46%) | 1.59 (1.15-2.18) | 1.64 (1.12-2.39) | 1.56 (1.05-2.33) |
| Cutaneous melanoma with sentinel node biopsy (C43, n = 1278) | |||||
| Never | 824 | 171 (83%) | Reference | Reference | Reference |
| Former | 239 | Not reached (75%) | 1.36 (1.06-1.76) | 1.34 (0.98-1.85) | 0.92 (0.66-1.28) |
| Current | 215 | Not reached (77%) | 1.38 (1.06-1.80) | 1.41 (1.02-1.96) | 1.29 (0.92-1.80) |
| Breast (C50, n = 9102) | |||||
| Never | 6079 | 170 (90%) | Reference | Reference | Reference |
| Former | 1522 | 174 (92%) | 0.80 (0.69-0.92) | 0.83 (0.69-1.01) | 0.97 (0.80-1.18) |
| Current | 1501 | 171 (89%) | 1.03 (0.91-1.17) | 1.06 (0.89-1.25) | 1.20 (1.02-1.44) |
| Uterine (C54-C55, n = 1289) | |||||
| Never | 969 | 169 (80%) | Reference | Reference | Reference |
| Former | 121 | 168 (76%) | 1.05 (0.75-1.47) | 0.97 (0.63-1.48) | 0.91 (0.60-1.25) |
| Current | 199 | Not reached (82%) | 0.80 (0.60-1.08) | 0.76 (0.53-1.10) | 0.95 (0.66-1.38) |
| Prostate (C61, n = 5357) | |||||
| Never | 2555 | 170 (93%) | Reference | Reference | Reference |
| Former | 1725 | 171 (90%) | 1.32 (1.16-1.50) | 1.39 (1.14-1.69) | 1.19 (0.98-1.46) |
| Current | 1077 | 173 (89%) | 1.45 (1.25-1.67) | 1.55 (1.25-1.93) | 1.76 (1.40-2.20) |
| Kidney (C64, n = 862) | |||||
| Never | 453 | 131 (75%) | Reference | Reference | Reference |
| Former | 188 | 160 (70%) | 1.17 (0.89-1.54) | 1.21 (0.86-1.70) | 1.04 (0.72-1.49) |
| Current | 221 | 133 (77%) | 1.15 (0.89-1.49) | 0.96 (0.69-1.34) | 1.10 (0.77-1.57) |
| Urinary tract (C65-C68, n = 1558) | |||||
| Never | 515 | 142 (68%) | Reference | Reference | Reference |
| Former | 514 | 79 (60%) | 1.50 (1.26-1.83) | 1.48 (1.17-1.86) | 1.39 (1.09-1.78) |
| Current | 529 | 101 (64%) | 1.28 (1.07-1.55) | 1.32 (1.04-1.67) | 1.55 (1.21-1.99) |
The definition of curative intent is presented in the Supplementary Material (available at https://doi.org/10.1016/j.esmorw.2026.100791). Statistically significant results are in bold.
CCI, Charlson comorbidity index; CI, confidence interval; CSS, cancer-specific survival; ECOG PS, Eastern Cooperative Oncology Group performance status; HR; hazard ratio; NSCLC, non-small-cell lung cancer; OS, overall survival.
Multivariable Cox regression adjusted for age >75 years, sex (where applicable), ECOG PS (0-1 versus 2-4 versus not available), CCI (0 versus 1 versus 2-10), alcohol abuse diagnosis, and nodal stage (node-negative versus node-positive versus not available).
Concerning former smokers, their OS varied strongly according to cancer site. The OS of former smokers was close to OS of never smokers with HNSCC and colorectal cancers, whereas for those with melanoma, gastro-oesophageal, pancreatic, and prostate cancer, the OS was close to that of current smokers. Former smokers who had breast cancer had the longest survival, whereas former smokers with uterine, kidney or urinary tract cancer had the shortest.
Baseline demographics differed according to smoking status (Supplementary Table S7, available at https://doi.org/10.1016/j.esmorw.2026.100791) among patients undergoing curative-intent treatment. Compared with never smokers, current smokers were younger, predominantly male, and more frequently presented with comorbidities, alcohol abuse, and low BMI. Additionally, poor ECOG PS together with current smoking was more frequently observed among patients with HNSCC or pancreatic cancer and node-positive disease was more frequently recorded in HNSCC and lung cancer.
Survival in patients treated with palliative intent
Treatment of 18 143 patients was classified as palliative at baseline, including 9064 patients actively treated and 9079 patients receiving BSC only. In the subgroup of actively treated patients, median OS was 27 months among never smokers (reference), 16 months among former (HR 1.40, 95% CI 1.32-1.50), and 11 months among current smokers (HR 1.79, 95% CI 1.68-1.90). These findings were confounded by a high lung cancer prevalence among current smokers (52% compared with 9%, Supplementary Table S8, available at https://doi.org/10.1016/j.esmorw.2026.100791).
Among patients allocated to the BSC-only group, with the caveat of including patients that may have received hormonal therapy or palliative radiotherapy only, the median OS was 16 months among never smokers (reference), 5 months among former smokers (HR 1.44, 95% CI 1.34-1.56), and 3 months among current smokers (HR 1.69, 95% CI 1.57-1.81).
Among the 7814 patients with ‘undefined’ treatment intent, median OS was 142 months among never smokers (reference), 102 months among former (HR 1.37, 95% CI 1.23-1.53), and 120 months among current smokers (HR 1.28, 95% CI 1.15-1.32).
Discussion
The present study evaluated the association between tobacco smoking status and CSS and OS after cancer diagnosis in a large, nearly complete population-based cohort of patients with solid cancers. Smoking was consistently associated with impaired OS and CSS across most cancer types and this association persisted after adjustment for major prognostic factors. Importantly, the survival disadvantage was also observed among patients treated with curative intent, suggesting that the survival disadvantage associated with smoking cannot be explained solely by differences in baseline characteristics or treatment allocation.
Impaired survival among smokers with cancer has been consistently reported in multiple malignancies. Previous epidemiological studies and meta-analyses have shown increased mortality among smokers with all major cancer types.5, 6, 7, 8, 9, 10, 11, 12,24 These studies have typically reported HRs for OS or CSS ranging between 1.2 and 2.2 among current smokers compared with never smokers, while the present study showed unadjusted HRs of 1.91 for OS and 2.09 for CSS. Across cancer types, the HRs were 1.15-3.98 for OS and 1.15-4.07 for CSS. These findings are broadly consistent with previously reported estimates,2,3,5,7,8,10, 11, 12,20,24 supporting the generalisability of our findings. Although smoking status was unavailable for 28% of patients, close to that in other studies,6,9,12,14 IPW analyses did not materially alter the estimates, suggesting limited selection bias related to the measured predictors of smoking-status availability. However, residual bias from other unmeasured factors, including socioeconomic status, frailty, and treatment adherence and intensity, cannot be excluded.
A key observation of the present study was that smoking-related survival differences persisted when analyses were restricted to patients receiving curative-intent treatment. This finding is particularly important because many previous studies have evaluated heterogeneous populations that include both curative- and palliative-intent treatment settings. Previous studies have proposed several biological explanations for this observation, including reduced radiosensitivity due to tumour hypoxia, impaired surgical recovery, altered tumour cell signalling, and drug metabolism associated with nicotine exposure.1, 2, 3,10,19
In the palliative setting, current smokers had shorter survival compared with never smokers, with former smokers showing intermediate outcomes. These findings should be interpreted with caution, as they are likely influenced by differences in cancer type distribution, particularly the higher proportion of lung cancer among smokers, a factor known to influence survival comparisons in smoking-related malignancies.1,2,9,19 Patients unfit for active cancer treatment are probably underrepresented in the current study. For example, only 80%-91% of lung cancer incidence was captured, a recognised problem in Finnish health care.25 Similar percentages were observed with pancreatic cancer and melanoma. The majority of superficial melanomas are treated outside university hospitals, where mainly extensive resections requiring sentinel lymph node biopsy are carried out,26 as reflected in melanoma survival in the present study.
Another noteworthy finding was that multivariable adjustment for age, sex, ECOG PS, CCI, alcohol abuse, treatment intent, and nodal status had a relatively modest impact on CSS estimates for most cancer types, except for breast cancer, where adjustment attenuated the association (HRs for adjusted CSS 1.80 for all and 1.14-2.14 across cancer types). This suggests that although smokers present with less favourable clinical characteristics, the survival disadvantage associated with smoking is only partly explained by these factors. Compared with never smokers, current smokers were diagnosed with cancer at a younger age but nevertheless had poorer survival outcomes. Similar observations have been reported in several cancer-specific studies and meta-analyses, in which smoking remained independently associated with impaired survival after adjustment for clinical prognostic factors.2,6,9,11,12,14 These findings support the hypothesis that smoking may directly influence tumour behaviour and treatment efficacy.1,2
A major strength of this study is the large, population-based cohort covering nearly one-fifth of the Finnish population and reflecting routine oncological care. The large sample size enabled robust subgroup analyses across multiple cancer types, integrating detailed clinical variables, treatment information, and nationwide mortality follow-up. The use of NLP to extract smoking status from electronic health records represents a scalable and validated approach20 to capture lifestyle-related data, addressing a key limitation of many other registry-based studies. Furthermore, the availability of treatment intent allowed separate evaluation of populations with curative- and palliative-intent treatment, which is rarely possible in large real-world datasets and provides important clinical context for interpreting survival differences.
The study has limitations. Although population-based within the university hospital catchment areas, the cohort was enriched with patients referred from other regions for centralised complex surgery and radiotherapy, including palliative radiotherapy, while some early-stage cancers, such as melanoma, that were treated outside of the university hospitals were underrepresented. Consequently, the cohort had a somewhat higher prevalence of metastatic disease and poorer survival than the nationwide cancer population. In addition, detailed information on smoking exposure such as pack-years or time from smoking cessation to cancer diagnosis was not available, preventing dose–response analyses. Because smoking status was based on the entire observational period, patients diagnosed earlier or surviving longer had more opportunity to quit smoking and subsequently be classified as former smokers, introducing potential immortal-time bias (Supplementary Table S5, available at https://doi.org/10.1016/j.esmorw.2026.100791). Changes in smoking behaviour after diagnosis may also introduce reverse-causation bias. Those managing to quit were classified by the NLP as former smokers, making this group heterogeneous and more challenging to interpret, as compared with never or current smokers. Most former smokers had quit smoking earlier, with only 18% around the time of cancer diagnosis, compared with cessation rates of 36%-43% that are usually reported.4,27 Finally, although several important clinical covariates were included in the multivariable models, residual confounding from unmeasured prognostic, socioeconomic, and lifestyle factors cannot be excluded; the observed HRs should therefore be interpreted as associations rather than causal effects.
Conclusions
In this large population-based cohort of patients with solid cancers, current or former smoking was consistently associated with impaired OS and CSS. The survival disadvantage persisted after adjustment for major prognostic factors and more importantly remained present among patients undergoing curative-intent treatment. These findings support the previous evidence that current smoking is associated with poorer cancer outcomes and highlight the importance of smoking assessment and cessation as part of comprehensive cancer care. The study also demonstrates the feasibility of identifying smoking status from electronic health records in population-based cancer research.
Informed consent
The authors have NOT obtained informed consent from participants or their legal representatives. For retrospective studies authorised by Findata, in which patients are not contacted, the requirements for informed consent and local Institutional Review Board approval are waived.
Ethical approval
This study was carried out in compliance with relevant laws, regulatory frameworks and guidelines where the research took place. Ethics committee approval was not required under relevant laws and institutional guidelines. The study was approved by the Finnish Social and Health Data Permit Authority Findata (THL/529/14.02.00/2022). Local administrative permissions were obtained from Turku (T132/2019) and Tampere (R19582) University Hospitals. Pseudonymised data from both hospitals were stored and analysed in an approved and secure processing environment.
Acknowledgements
This study was financially supported by Juho Vainio foundation and State Competitive Research Funding of Turku and Tampere University Hospitals. We thank all personnel at Findata and clinical informatics teams at Turku and Tampere University Hospitals for their assistance, and Adelaide Lönnberg (MapleMountain Editing) for the language revision.
Funding
This work was supported by the Juho Vainio Foundation [E. Heervä, 2021 and 2025], State Competitive Research Funding of Turku [E. Heervä, 2021 and 2025], and Tampere University Hospital [P. Österlund, 2023-2026], grant numbers not applicable.
Disclosure
The authors have declared no conflicts of interest.
Data availability
The pretrained ULMFiT language model is available at https://github.com/AnttiKarlsson/finnish_ulmfit. The pretrained Finnish BERT model is available at http://turkunlp.org/FinBERT/. The clinical data can be requested from findata.fi/en.
Supplementary data
References
- 1.Bergman M., Fountoukidis G., Smith D., Ahlgren J., Lambe M., Valachis A. Effect of smoking on treatment efficacy and toxicity in patients with cancer: a systematic review and meta-analysis. Cancers (Basel) 2022;14(17):4117. doi: 10.3390/cancers14174117. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Schaefers C., Seidel C., Bokemeyer F., Bokemeyer C. The prognostic impact of the smoking status of cancer patients receiving systemic treatment, radiation therapy, and surgery: a systematic review and meta-analysis. Eur J Cancer. 2022;172:130–137. doi: 10.1016/j.ejca.2022.05.027. [DOI] [PubMed] [Google Scholar]
- 3.Warren G.W., Kasza K.A., Reid M.E., Cummings K.M., Marshall J.R. Smoking at diagnosis and survival in cancer patients. Int J Cancer. 2013;132(2):401–410. doi: 10.1002/ijc.27617. [DOI] [PubMed] [Google Scholar]
- 4.Cinciripini P.M., Kypriotakis G., Blalock J.A., et al. Survival outcomes of an early intervention smoking cessation treatment after a cancer diagnosis. JAMA Oncol. 2024;10(12):1689–1696. doi: 10.1001/jamaoncol.2024.4890. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Bérubé S., Lemieux J., Moore L., Maunsell E., Brisson J. Smoking at time of diagnosis and breast cancer-specific survival: new findings and systematic review with meta-analysis. Breast Cancer Res. 2014;16(2) doi: 10.1186/bcr3646. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Hamajima N., Hirose K., Tajima K., et al. Alcohol, tobacco and breast cancer–collaborative reanalysis of individual data from 53 epidemiological studies, including 58,515 women with breast cancer and 95,067 women without the disease. Br J Cancer. 2002;87(11):1234–1245. doi: 10.1038/sj.bjc.6600596. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Walter V., Jansen L., Hoffmeister M., Brenner H. Smoking and survival of colorectal cancer patients: systematic review and meta-analysis. Ann Oncol. 2014;25(8):1517–1525. doi: 10.1093/annonc/mdu040. [DOI] [PubMed] [Google Scholar]
- 8.Foerster B., Pozo C., Abufaraj M., et al. Association of smoking status with recurrence, metastasis, and mortality among patients with localized prostate cancer undergoing prostatectomy or radiotherapy: a systematic review and meta-analysis. JAMA Oncol. 2018;4(7):953–961. doi: 10.1001/jamaoncol.2018.1071. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Fares A.F., Li Y., Jiang M., et al. Association between duration of smoking abstinence before non-small-cell lung cancer diagnosis and survival: a retrospective, pooled analysis of cohort studies. Lancet Public Health. 2023;8(9):e691–e700. doi: 10.1016/S2468-2667(23)00131-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Smith J., Nastasi D., Tso R., Vangaveti V., Renison B., Chilkuri M. The effects of continued smoking in head and neck cancer patients treated with radiotherapy: a systematic review and meta-analysis. Radiother Oncol. 2019;135:51–57. doi: 10.1016/j.radonc.2019.02.021. [DOI] [PubMed] [Google Scholar]
- 11.Cacciamani G.E., Ghodoussipour S., Mari A., et al. Association between smoking exposure, neoadjuvant chemotherapy response and survival outcomes following radical cystectomy: systematic review and meta-analysis. J Urol. 2020;204(4):649–660. doi: 10.1097/JU.0000000000000813. [DOI] [PubMed] [Google Scholar]
- 12.Friedman E.B., Williams G.J., Lo S.N., Thompson J.F. Outcomes for smokers who develop melanoma: a systematic review and meta-analysis. EClinicalMedicine. 2024;77 doi: 10.1016/j.eclinm.2024.102872. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Xu Y., Qi Y., Zhang J., et al. The impact of smoking on survival in renal cell carcinoma: a systematic review and meta-analysis. Tumour Biol. 2014;35(7):6633–6640. doi: 10.1007/s13277-014-1862-8. [DOI] [PubMed] [Google Scholar]
- 14.Lindemann K., Vatten L.J., Ellstrøm-Engh M., Eskild A. Body mass, diabetes and smoking, and endometrial cancer risk: a follow-up study. Br J Cancer. 2008;98(9):1582–1585. doi: 10.1038/sj.bjc.6604313. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Terry P.D., Rohan T.E., Franceschi S., Weiderpass E. Cigarette smoking and the risk of endometrial cancer. Lancet Oncol. 2002;3(8):470–480. doi: 10.1016/s1470-2045(02)00816-1. [DOI] [PubMed] [Google Scholar]
- 16.Tabuchi T., Ito Y., Ioka A., Nakayama T., Miyashiro I., Tsukuma H. Tobacco smoking and the risk of subsequent primary cancer among cancer survivors: a retrospective cohort study. Ann Oncol. 2013;24(10):2699–2704. doi: 10.1093/annonc/mdt279. [DOI] [PubMed] [Google Scholar]
- 17.Li Q., Zhang G., Li X., et al. Risk of cardiovascular disease among cancer survivors: systematic review and meta-analysis. EClinicalMedicine. 2025;84 doi: 10.1016/j.eclinm.2025.103274. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Mills E., Eyawo O., Lockhart I., Kelly S., Wu P., Ebbert J.O. Smoking cessation reduces postoperative complications: a systematic review and meta-analysis. Am J Med. 2011;124(2):144–154.e8. doi: 10.1016/j.amjmed.2010.09.013. [DOI] [PubMed] [Google Scholar]
- 19.Li X., Huang C., Xie X., et al. The impact of smoking status on the progression-free survival of non-small cell lung cancer patients receiving molecularly target therapy or immunotherapy versus chemotherapy: a meta-analysis. J Clin Pharm Ther. 2021;46(2):256–266. doi: 10.1111/jcpt.13309. [DOI] [PubMed] [Google Scholar]
- 20.Karlsson A., Ellonen A., Irjala H., et al. Impact of deep learning-determined smoking status on mortality of cancer patients: never too late to quit. ESMO Open. 2021;6(3) doi: 10.1016/j.esmoop.2021.100175. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Lahti R.A., Penttilä A. Cause-of-death query in validation of death certification by expert panel; effects on mortality statistics in Finland, 1995. Forensic Sci Int. 2003;131(2-3):113–124. doi: 10.1016/s0379-0738(02)00418-8. [DOI] [PubMed] [Google Scholar]
- 22.Kilpeläinen T.P., Mäkinen T., Karhunen P.J., et al. Estimating bias in causes of death ascertainment in the Finnish Randomized Study of Screening for Prostate Cancer. Cancer Epidemiol. 2016;45:1–5. doi: 10.1016/j.canep.2016.08.022. [DOI] [PubMed] [Google Scholar]
- 23.Borras J.M., Barton M., Grau C., et al. The impact of cancer incidence and stage on optimal utilization of radiotherapy: methodology of a population based analysis by the ESTRO-HERO project. Radiother Oncol. 2015;116(1):45–50. doi: 10.1016/j.radonc.2015.04.021. [DOI] [PubMed] [Google Scholar]
- 24.Fahey P.P., Mallitt K.-A., Astell-Burt T., Stone G., Whiteman D.C. Impact of pre-diagnosis behavior on risk of death from esophageal cancer: a systematic review and meta-analysis. Cancer Causes Control. 2015;26(10):1365–1373. doi: 10.1007/s10552-015-0635-z. [DOI] [PubMed] [Google Scholar]
- 25.Paappanen V., Tikka J., Jukkola A., et al. Outcome of lung cancer surgery and proportion of lung cancer patients eligible for surgery in five Finnish hospitals in 2018, real world study. Acta Oncol. 2025;64:909–916. doi: 10.2340/1651-226X.2025.43398. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Mattila K., Vihinen H., Karlsson A., Minn H., Vihinen P., Heervä E. Smoking is an independent marker of poor prognosis in cutaneous melanoma. Acta Derm Venereol. 2023;103 doi: 10.2340/actadv.v103.3209. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Caini S., Gorini G., Del Riccio M., et al. Upgrading your best chances: postdiagnosis smoking cessation boosts life expectancy of patients with cancer–a systematic review and meta-analysis. Tob Control. 2025 doi: 10.1136/tc-2024-058873. 6:tc-2024-058873. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The pretrained ULMFiT language model is available at https://github.com/AnttiKarlsson/finnish_ulmfit. The pretrained Finnish BERT model is available at http://turkunlp.org/FinBERT/. The clinical data can be requested from findata.fi/en.