A machine learning tool for identifying non-metastatic colorectal cancer in primary care.

BACKGROUND: Primary health care (PHC) is often the first point of contact when diagnosing colorectal cancer (CRC). Human limitations in processing large amounts of information warrant the use of machine learning as a diagnostic prediction tool for CRC. AIM: To develop a predictive model for identifying non-metastatic CRC (NMCRC) among PHC patients using diagnostic data analysed with machine learning. DESIGN AND SETTING: A case-control study containing data on PHC visits for 542 patients >18 years old diagnosed with NMCRC in the Västra Götaland Region, Sweden, during 2011, and 2,139 matched controls. METHOD: Stochastic gradient boosting (SGB) was used to construct a model for predicting the presence of NMCRC based on diagnostic codes from PHC consultations during the year before the date of cancer diagnosis and the total number of consultations. Variables with a normalised relative influence (NRI) >1% were considered having an important contribution to the model. Risks of having NMCRC were calculated using odds ratios of marginal effects. RESULTS: Of the 361 variables used as predictors in the stochastic gradient boosting model, 184 had non-zero influence, with 16 variables having NRI >1% and a combined NRI of 63.3%. Variables representing anaemia and bleeding had a combined NRI of 27.6%. The model had a sensitivity of 73.3% and a specificity of 83.5%. Change in bowel habit had the highest odds ratios of marginal effects at 28.8. CONCLUSION: Machine learning is useful for identifying variables of importance for predicting NMCRC in PHC. Malignant diagnoses may be hidden behind benign symptoms such as haemorrhoids.

Validation of a diagnostic prediction tool for colorectal cancer: a case-control replication study.

BACKGROUND: Early detection of colorectal cancer (CRC) is crucial for survival. Primary care, the first point of contact in most cases, needs supportive risk assessment tools. We aimed to replicate the Swedish Colorectal Cancer Risk Assessment Tool (SCCRAT) for non-metastatic CRC in primary care and examine if risk factor patterns depend on sex and age. METHODS: 2,920 adults diagnosed with non-metastatic CRC during the years 2015-2019 after having visited a general practitioner the year before the diagnosis were selected from the Swedish Cancer Register and matched with 11,628 controls, using the same inclusion criteria except for the CRC diagnosis. Diagnostic codes from primary care consultations were collected from a regional health care database. Positive predictive values (PPVs) were estimated for the same 5 symptoms and combinations thereof as in the baseline study. RESULTS: The results for patients aged ≥50 years old in the present study were consistent with the results of the SCCRAT study. All symptoms and combinations thereof with a PPV >5% in the present study had a PPV >5% in the baseline study. The combination of bleeding with abdominal pain (PPV 9.9%) and bleeding with change in bowel habit (PPV 7.8%) were the highest observed PPVs in both studies. Similar risk patterns were seen for all ages and when men and women were studied separately. CONCLUSION: This external validation of the SCCRAT for non-metastatic CRC in primary care replicated the baseline study successfully and identified patients at high risk for CRC.

Clinical features of patients with non-metastatic lung cancer in primary care: a case-control study.

BACKGROUND: Lung cancer (LC) kills more people than any other cancer globally, mainly due to the late stage of diagnosis. AIM: To identify and quantify the prediagnostic features of non-metastatic lung cancer (nMLC) and to compare the clinical features in GPs' chest X-ray referral letters with the clinical features (expressed as diagnostic codes) in medical records. DESIGN & SETTING: A population-based case-control study was conducted using diagnostic codes from national and regional healthcare databases in Sweden. METHOD: In total, 373 patients diagnosed with LC in 2011 (of which 132 had nMLC) and 1472 controls were selected from the Swedish Cancer Register (SCR) and regional healthcare database, respectively. Diagnostic codes registered in medical records from primary care consultations in the year before LC diagnosis were collected from the regional healthcare database. Odds ratios (OR) were calculated for variables associated with nMLC. The GPs' referral letters for chest X- ray were retrieved from the regional repository for radiology. RESULTS: Clinical features with the highest OR were vitamin B12 deficiency anaemia (OR 6.7, 95% confidence interval [CI] = 1.6 to 27.9), dyspnoea (OR 5.0, 95% CI = 2.0 to 12.7), and chronic bronchitis (OR 5.0, 95% CI = 1.3 to 18.6). Clinical features that were GPs' reasons for requesting chest X-ray were almost three times more frequent in referral letters compared to the corresponding diagnostic codes in the medical records. CONCLUSION: Patients with nMLC could not be identified by symptoms. The clinical features in referral letters for X-ray were more frequent than corresponding diagnostic codes from medical records.

Diagnostic profile characteristics of cancer patients with frequent consultations in primary care before diagnosis: a case-control study.

Background: Many patients with common cancers are late diagnosed. Objectives: Identify consultation profiles and clinical features in patients with the seven most common cancers, who had consulted a general practitioner (GP) frequently before their cancer diagnosis. Methods: A case-control study was conducted in Region Västra Götaland, Sweden. A total of 2570 patients, diagnosed in 2011 with prostate, breast, colorectal, lung, gynaecological and skin cancers including malignant melanoma, and 9424 controls were selected from the Swedish Cancer Register and a regional health care database. Diagnostic codes [International Statistical Classification of Diseases and Related Health Problems, 10th revision (ICD-10)] from primary care for patients with ≥4 GP consultations registered in the year before cancer diagnosis were collected. Likelihood ratios (LRs) were calculated for variables associated with the different cancers. Results: Fifty-six percent of the patients had consulted a GP four or more times in the year before cancer diagnosis. Alarm symptoms or signs represented 60% of the codes with the highest LR, but only 40% of the 10 most prevalent codes. Breast lump had the highest LR, 11.9 [95% confidence interval (CI) 8.0-17.8]; abnormalities of plasma proteins had an LR of 5.0 (95% CI 3.0-8.2) and abnormal serum enzyme levels had an LR of 4.6 (95% CI 3.6-5.9). Early clinical features associated with cancer had been registered already at the first two GP consultations. Conclusion: One out of six clinical features associated with cancer were presented by cancer patients with four or more pre-referral consultations already at the two first consultations. These early clinical features that were focal and had benign characteristics might have been missed diagnostic opportunities.

Identification of patients with non-metastatic colorectal cancer in primary care: a case-control study.

BACKGROUND: Colorectal cancer is the third most common cancer worldwide and second most common in Europe. Despite screening, it is often diagnosed at an unfavourable stage. AIM: To identify and quantify features of non-metastatic colorectal cancer in primary care to enable earlier diagnosis by GPs. DESIGN AND SETTING: A case-control study was conducted using diagnostic codes from national and regional healthcare databases in Sweden. METHOD: A total of 542 patients diagnosed with non-metastatic colorectal cancer in 2011 and 2139 matched controls were selected from the Swedish Cancer Register (SCR) and a regional healthcare database respectively. All diagnostic codes (according to ICD-10) from primary care consultations registered the year before the date of cancer diagnosis (according to the SCR) were collected from the regional database. Odds ratios were calculated for variables independently associated with non-metastatic colorectal cancer using multivariable conditional logistic regressions. Positive predictive values (PPVs) of these variables were calculated, both individually and in combination with each other. RESULTS: Five features were associated with colorectal cancer before diagnosis: bleeding, including rectal bleeding, melaena, and gastrointestinal bleeding (PPV 3.9%, 95% confidence interval [CI] = 2.3 to 6.3); anaemia (PPV 1.4%, 95% CI = 1.1 to 1.8); change in bowel habit (PPV 1.1%, 95% CI = 0.9 to 1.5; abdominal pain (PPV 0.9%, 95% CI = 0.7 to 1.1); and weight loss (PPV 1.0%, 95% CI = 0.3 to 3.0); all P-value <0.05. The combination of bleeding and change in bowel habit had a PPV of 13.7% (95% CI = 2.1 to 54.4); for bleeding combined with abdominal pain this was 12.2% (95% CI = 1.8 to 51.2). A risk assessment tool for non-metastatic colorectal cancer was designed. CONCLUSION: Bleeding combined with either diarrhoea, constipation, change in bowel habit, or abdominal pain are the most powerful predictors of non-metastatic colorectal cancer and should result in prompt referral for colorectal investigation.

Increased consultation frequency in primary care, a risk marker for cancer: a case-control study.

OBJECTIVE: To identify early diagnostic profiles such as diagnostic codes and consultation patterns of cancer patients in primary care one year prior to cancer diagnosis. DESIGN: Total population-based case-control study. SETTING AND SUBJECTS: 4562 cancer patients and 17,979 controls matched by age, sex, and primary care unit. Data were collected from the Swedish Cancer Register and the Regional Healthcare Database. METHOD: We identified cancer patients in the Västra Götaland Region of Sweden diagnosed in 2011 with prostate, breast, colorectal, lung, gynaecological, and skin cancers including malignant melanoma. We studied the symptoms and diagnoses identified by diagnostic codes during a diagnostic interval of 12 months before the cancer diagnosis. MAIN OUTCOME MEASURES: Consultation frequency, symptom density by cancer type, prevalence and odds ratios (OR) for the diagnostic codes in the cancer population as a whole. RESULTS: The diagnostic codes with the highest OR were unspecified lump in breast, neoplasm of uncertain behaviour, and abnormal serum enzyme levels. The codes with the highest prevalence were hyperplasia of prostate, other skin changes and abdominal and pelvic pain. The frequency of diagnostic codes and consultations in primary care rose in tandem 50 days before diagnosis for breast and gynaecological cancer, 60 days for malignant melanoma and skin cancer, 80 days for prostate cancer and 100 days for colorectal and lung cancer. CONCLUSION: Eighty-seven percent of patients with the most common cancers consulted a general practitioner (GP) a year before their diagnosis. An increase in consultation frequency and presentation of any symptom should raise the GP's suspicion of cancer. Key points Knowledge about the prevalence of early symptoms and other clinical signs in cancer patients in primary care remains insufficient. • Eighty-seven percent of the patients with the seven most common cancers consulted a general practitioner 12 months prior to cancer diagnosis. • Both the frequency of consultation and the number of symptoms and diseases expressed in diagnostic codes rose in tandem 50-100 days before the cancer diagnosis. • Unless it is caused by a previously known disease, an increased consultation rate for any symptom should result in a swift investigation or referral from primary care to confirm or exclude cancer.