In a recently published article in Nature medicineThe researchers applied artificial intelligence (AI) methods to real-world longitudinal clinical data to design surveillance programs for the early detection of patients at risk of developing one of the most serious diseases, pancreatic cancer.
Stady: Deep learning algorithm for predicting pancreatic cancer risk from disease pathways. Image credit: Chinnapong/Shutterstock.com
The incidence of pancreatic cancer is increasing, making it a leading cause of cancer-related deaths worldwide. Pancreatic cancer is difficult to diagnose because its risk factors are not understood.
Late detection in advanced or distant metastatic stages hinders treatment, making patient survival uncommon. Only two to nine percent of these patients survive to five years of age.
While age is a recognized risk factor for pancreatic cancer, population-wide screening based on age is impractical due to the high cost of clinical testing, which also leads to false positive results.
In addition, data on family history or genetic risk factors for the general population is often not available. Thus, there is an urgent need to develop affordable surveillance programs for early detection of pancreatic cancer in the general population.
In this study, the researchers used realistic longitudinal clinical records of large numbers of patients to identify a very small number of patients at high risk of developing pancreatic cancer.
They exploited recently developed machine learning (ML) methods using patient records from the Danish National Patient Registry (DNPR) and, in turn, from the company’s data repository of the United States Veterans Affairs (US-VA).
The former included data on 8.6 million patients enrolled between 1977 and 2018, which equates to 24,000 cases of pancreatic cancer, while the latter had clinical data on 3 million patients with 3,900 cases of pancreatic cancer.
The team trained and tested a variety of ML models on sequences of disease codes in DNPR and US-VA clinical registries and tested prediction of cancer incidence over additional time periods called CancerRiskNet.
In building the predictive models, the team used the three-letter International Classification of Diseases (ICD) diagnostic codes and defined “Pancreatic cancer patientsAs patients had at least one code below C25, indicating malignancy of the pancreas.
The accuracy of the cancer diagnostic codes was ~98%. Finally, the researchers determined which diagnoses in the patient’s history of the diagnostic codes were most informative of cancer risk to suggest an optimal surveillance programme.
Furthermore, the researchers evaluated the prediction performance of different DNPR-trained models using the area under receiver operating characteristic (AUROC) and relative hazards (RR) curves. In addition, they reported ML-derived RR scores for patients with cancer in the high-risk group.
All previous studies using realistic clinical records to predict the risk of pancreatic cancer have brought encouraging results but have not used disease history chronology to extract time-series longitudinal features. In this study, they evaluated non-sequencing models on the DNPR dataset.
Overall, the time-series model, Transformer, had the best performance in predicting the incidence of cancer within 36 months of the evaluation date, with an AUROC of 0.879, closely followed by the GRU with an AUROC of 0.852.
The RR of this model at operational point defined by n = 1000 highest-risk patients out of 1 million patients was 104.7.
The performance of the word bag model and the MLP model for predicting the incidence of cancer within 36 months in terms of AUROC was 0.807 and 0.845, respectively. However, compared to Transformer, the RRs for word bags and MLP were significantly lower (104.7 vs 2.1 and 26.6).
Excluding the data, that is, excluding input disease diagnoses from the past three, six, and 12 months prior to the diagnosis of pancreatic cancer, decreased the performance of best models AUROCs of 0.879 to AUROCs of 0.843, 0.829, and 0.827 for three/six/12 months.
This analysis indicated that the ML model trained on data from both sources had a positive predictive value (PPV) of 0.32 for the 12-month prediction period. Therefore, about 320 patients eventually developed pancreatic cancer.
While clinicians may have identified some cases based on recognized risk factors for pancreatic cancer, eg, chronic pancreatitis, only a small fraction of them, approximately 70, are still newly identified at a conservative estimate.
Despite the use of common ICD disease codes and ipsilateral cancer survival, cross-application of DNPR data to US-VA data reduced the performance of the ML models, thus increasing the need for independent model training across geographic regions to achieve optimal performance of the regional model.
However, the ideal scenario for multi-institutional collaboration to achieve a globally relevant set of prediction rules would require standardized learning across different healthcare systems.
The prediction accuracy of the ML-based models described in this study could improve with access to data other than disease codes, for example, written notes in clinical notes, laboratory results and genetic profiles of more people or health-related information from their wearable devices.
Next, the clinical implementation of early diagnosis of pancreatic cancer will require the identification of high-risk patients.
Because those at highest risk are a smaller subset of a large computationally screened population, expensive and duplicate clinical screening and intervention programs will be limited to a small number of patients.
However, AI in clinical records from the real world can shift the focus from late-stage treatment to early-stage cancer treatment, which in turn will greatly improve the quality of life of all patients while increasing the cost-benefit ratio. Cancer care.