Pancreatic cancer symptom trajectories from Danish registry data and free text in electronic health records
Curation statements for this article:-
Curated by eLife
eLife assessment
This study presents valuable findings on the symptoms and disease trajectories preceding a diagnosis of pancreatic cancer in Denmark. The evidence supporting the claims of the authors is solid, although an error analysis of the text mining evaluation results and a discussion on how the findings can be applied in practice would strengthen the study. The work will be of interest to public health researchers and clinicians working on pancreatic cancer.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
Abstract
Pancreatic cancer is one of the deadliest cancer types with poor treatment options. Better detection of early symptoms and relevant disease correlations could improve pancreatic cancer prognosis. In this retrospective study, we used symptom and disease codes (ICD-10) from the Danish National Patient Registry (NPR) encompassing 6.9 million patients from 1994 to 2018,, of whom 23,592 were diagnosed with pancreatic cancer. The Danish cancer registry included 18,523 of these patients. To complement and compare the registry diagnosis codes with deeper clinical data, we used a text mining approach to extract symptoms from free text clinical notes in electronic health records (3078 pancreatic cancer patients and 30,780 controls). We used both data sources to generate and compare symptom disease trajectories to uncover temporal patterns of symptoms prior to pancreatic cancer diagnosis for the same patients. We show that the text mining of the clinical notes was able to complement the registry-based symptoms by capturing more symptoms prior to pancreatic cancer diagnosis. For example, ‘Blood pressure reading without diagnosis’, ‘Abnormalities of heartbeat’, and ‘Intestinal obstruction’ were not found for the registry-based analysis. Chaining symptoms together in trajectories identified two groups of patients with lower median survival (<90 days) following the trajectories ‘Cough→Jaundice→Intestinal obstruction’ and ‘Pain→Jaundice→Abnormal results of function studies’. These results provide a comprehensive comparison of the two types of pancreatic cancer symptom trajectories, which in combination can leverage the full potential of the health data and ultimately provide a fuller picture for detection of early risk factors for pancreatic cancer.
Article activity feed
-
-
Author Response
Reviewer #2 (Public Review):
This manuscript reports on an important study that aims to identify symptom trajectories for the early detection of pancreatic cancer. The study's findings are based on the analysis of two complementary data sources: structured data obtained from the Danish National Patient Registry and unstructured information extracted from the free-text sections of patient notes. The researchers successfully identified various symptoms and disease trajectories that are strongly associated with pancreatic cancer, with compelling evidence from both data sources. Additionally, the study provides a detailed comparison and contrast of the results obtained from each data source, adding valuable insights into the strengths and limitations of each method.
Strengths:
The work is well motivated by the urgent …
Author Response
Reviewer #2 (Public Review):
This manuscript reports on an important study that aims to identify symptom trajectories for the early detection of pancreatic cancer. The study's findings are based on the analysis of two complementary data sources: structured data obtained from the Danish National Patient Registry and unstructured information extracted from the free-text sections of patient notes. The researchers successfully identified various symptoms and disease trajectories that are strongly associated with pancreatic cancer, with compelling evidence from both data sources. Additionally, the study provides a detailed comparison and contrast of the results obtained from each data source, adding valuable insights into the strengths and limitations of each method.
Strengths:
The work is well motivated by the urgent need for early detection of pancreatic cancer, which is often difficult due to the lack of effective (computational) methods. The manuscript is generally well-written and includes relevant studies, providing a comprehensive overview of the current state of the field.
One of the unique contributions of this work is its use of both structured registry data and unstructured clinical notes to leverage complementary information. This approach enables a more nuanced and comprehensive understanding of the disease symptom trajectories, which is critical for improving early disease diagnosis and prognosis.
The methodology employed in this study is sound and robust, and the authors have candidly discussed its limitations. The results are significant and highlight previously unknown insights into symptom disease trajectories, which have important implications for the management of pancreatic cancer.
Overall, this is a well-designed and executed study that makes an important contribution to the field of cancer/informatics research, and it should be of great interest to both researchers and clinicians.
Weaknesses:
To complement the results in Figure 1, I'd also suggest that the authors compile a list of the most common (known) symptoms of pancreatic cancer as a reference. In other words, not only can you compare results found from the two sources but also compare them with existing knowledge. This is something you discussed partly in lines 245 but including this early as part of the results in Figure 1 would be more informative.
We agree that this would be informative to include into the Venn diagram. Hence, we have created a list of the most established and well-known symptoms of pancreatic cancer (Supplementary table S1) and converted these to the comparable ICD-10 level that we also use for the text mining and registry counts in Fig. 1. We have included the Venn diagram as Supplementary Figure S1.
In terms of the text mining evaluation results, providing information on recall errors would be beneficial to better understand the performance of the method. Additionally, line 144 mentions 53 words, but it is still not clear to me what these words refer to. Could you please clarify this point or provide more context?
We have added sensitivity/recall measures on the text mining procedure and furthermore added two references in the Discussion of the Tagcorpus program which was used for text mining the clinical notes. These references also mention similar sensitivities for the studies. The 53 words are false positives and we have clarified why these have been captured as false positives by the Tagcorpus (negations).
The disparities between Figure 2A and 2B are noteworthy, from very different initial symptoms to the proportion of short median survival dates (<=90 days), with much more pronounced differences than those observed in Figure 1 comparing two data sources. The highlighted trajectories are almost completely different. Should this be expected? I was hoping to see at least some overlap between the two results.
After updating the case population (via the cancer registry) and showing only symptoms trajectories in this revised version, we can clearly see that the trajectories are more similar. This gives an indication that the methods pick up on similar pancreatic-cancer symptoms, but there are also differences that show how each data type can complement the other, such as the text-mined trajectories being able to pick up longer symptom trajectories prior to the cancer.
All trajectories shown in Figure 2 include three symptoms. Is this by design? Could there be meaningful trajectories with different numbers of symptoms (e.g. 4 or more)?
We agree and have added the significant length 4 trajectories (for the registry data) as supplementary figure S2. No trajectories with length 5 or higher were found in the registry-based analysis. No length 4 (or higher) trajectories were found for the text-mined patients (presumably due to the data set size).
Considering those patients with both clinical notes and registry data, it may be beneficial to merge their symptoms to generate more informative trajectories.
This could be interesting but is out of scope for this paper. Here we would like to stress the proof-of-concept that the two data types can complement each other. The next steps would be to generate these multimodal trajectories to for example test if they are predictive of pancreatic cancer. Nonetheless, we acknowledge the significance of this perspective and have incorporated it into the Discussion section of the manuscript.
Given that results from two sources are being compared in Figures 1 and 2, have you considered calculating the top 20 most significant symptoms from the registry data as well?
We have done this and added them to Supplementary figure S3.
While there is a discussion related to cardiovascular diseases, I noticed no mention of cataracts or gonarthrosis, which were found to be prevalent among patients with short survival in Figure 2.
Since we now only include symptoms trajectories in the Results, we have chosen to not include these results in the Discussion for the final version of the manuscript. However, the diagnosis-wide trajectories are included in the Supplementary figure S2. Cataract and gonarthrosis have still been found significant in the results even though they are not shown in the Supplementary figure due to its visualization threshold of min. 400 patients per trajectory.
Ultimately, the goal of this research is to improve the early detection and prognosis of pancreatic cancer, thus it is important to discuss how the findings of this work could be applied in practice towards this goal (e.g. used by disease prediction algorithms?)
We agree that this is very important and have added a small section on this in the Discussion. We have also cited a recent publication using deep learning algorithms to predict pancreatic cancer based solely on registry data (Placido et al. 2023).
-
eLife assessment
This study presents valuable findings on the symptoms and disease trajectories preceding a diagnosis of pancreatic cancer in Denmark. The evidence supporting the claims of the authors is solid, although an error analysis of the text mining evaluation results and a discussion on how the findings can be applied in practice would strengthen the study. The work will be of interest to public health researchers and clinicians working on pancreatic cancer.
-
Reviewer #1 (Public Review):
The study by Hjaltelin, Novitski, and colleagues analyses clinical records of people with pancreatic cancer in the 5 years prior to their diagnosis, aiming to determine patterns of symptoms and disease trajectories that precede the pancreatic cancer diagnosis. The authors use established methodology to identify temporal disease patterns from the Danish National Patient Registry, covering >22,700 patients with pancreatic cancer and 8.1 million controls. They also apply a text-mining approach to extract potential symptoms from free-text clinical notes of a subset of individuals (>4,400 people with pancreatic cancer and >44,000 controls).
The large datasets used in this study present a very clear strength, and the results are presented quite clearly.
Weaknesses of the study include the relatively low …
Reviewer #1 (Public Review):
The study by Hjaltelin, Novitski, and colleagues analyses clinical records of people with pancreatic cancer in the 5 years prior to their diagnosis, aiming to determine patterns of symptoms and disease trajectories that precede the pancreatic cancer diagnosis. The authors use established methodology to identify temporal disease patterns from the Danish National Patient Registry, covering >22,700 patients with pancreatic cancer and 8.1 million controls. They also apply a text-mining approach to extract potential symptoms from free-text clinical notes of a subset of individuals (>4,400 people with pancreatic cancer and >44,000 controls).
The large datasets used in this study present a very clear strength, and the results are presented quite clearly.
Weaknesses of the study include the relatively low sensitivity of the text-mining approach to identify symptoms (83.4%) and the comparison of findings from datasets including different individuals (rather than a comparison of findings based on free-text entries and diagnosis codes from specific entries for the same patients). It is also not clear which proportion of patients is captured by the symptom and disease trajectories catalogued in this work. The different average survival times associated with different trajectories are very interesting, and it would be helpful to examine whether these are due to differences in cancer stage at diagnosis.
-
Reviewer #2 (Public Review):
This manuscript reports on an important study that aims to identify symptom trajectories for the early detection of pancreatic cancer. The study's findings are based on the analysis of two complementary data sources: structured data obtained from the Danish National Patient Registry and unstructured information extracted from the free-text sections of patient notes. The researchers successfully identified various symptoms and disease trajectories that are strongly associated with pancreatic cancer, with compelling evidence from both data sources. Additionally, the study provides a detailed comparison and contrast of the results obtained from each data source, adding valuable insights into the strengths and limitations of each method.
Strengths:
The work is well motivated by the urgent need for early …
Reviewer #2 (Public Review):
This manuscript reports on an important study that aims to identify symptom trajectories for the early detection of pancreatic cancer. The study's findings are based on the analysis of two complementary data sources: structured data obtained from the Danish National Patient Registry and unstructured information extracted from the free-text sections of patient notes. The researchers successfully identified various symptoms and disease trajectories that are strongly associated with pancreatic cancer, with compelling evidence from both data sources. Additionally, the study provides a detailed comparison and contrast of the results obtained from each data source, adding valuable insights into the strengths and limitations of each method.
Strengths:
The work is well motivated by the urgent need for early detection of pancreatic cancer, which is often difficult due to the lack of effective (computational) methods. The manuscript is generally well-written and includes relevant studies, providing a comprehensive overview of the current state of the field.
One of the unique contributions of this work is its use of both structured registry data and unstructured clinical notes to leverage complementary information. This approach enables a more nuanced and comprehensive understanding of the disease symptom trajectories, which is critical for improving early disease diagnosis and prognosis.
The methodology employed in this study is sound and robust, and the authors have candidly discussed its limitations. The results are significant and highlight previously unknown insights into symptom disease trajectories, which have important implications for the management of pancreatic cancer.
Overall, this is a well-designed and executed study that makes an important contribution to the field of cancer/informatics research, and it should be of great interest to both researchers and clinicians.
Weaknesses:
To complement the results in Figure 1, I'd also suggest that the authors compile a list of the most common (known) symptoms of pancreatic cancer as a reference. In other words, not only can you compare results found from the two sources but also compare them with existing knowledge. This is something you discussed partly in lines 245 but including this early as part of the results in Figure 1 would be more informative.
In terms of the text mining evaluation results, providing information on recall errors would be beneficial to better understand the performance of the method. Additionally, line 144 mentions 53 words, but it is still not clear to me what these words refer to. Could you please clarify this point or provide more context?
The disparities between Figure 2A and 2B are noteworthy, from very different initial symptoms to the proportion of short median survival dates (<=90 days), with much more pronounced differences than those observed in Figure 1 comparing two data sources. The highlighted trajectories are almost completely different. Should this be expected? I was hoping to see at least some overlap between the two results.
All trajectories shown in Figure 2 include three symptoms. Is this by design? Could there be meaningful trajectories with different numbers of symptoms (e.g. 4 or more)?
Considering those patients with both clinical notes and registry data, it may be beneficial to merge their symptoms to generate more informative trajectories.
Given that results from two sources are being compared in Figures 1 and 2, have you considered calculating the top 20 most significant symptoms from the registry data as well?
While there is a discussion related to cardiovascular diseases, I noticed no mention of cataracts or gonarthrosis, which were found to be prevalent among patients with short survival in Figure 2.
Ultimately, the goal of this research is to improve the early detection and prognosis of pancreatic cancer, thus it is important to discuss how the findings of this work could be applied in practice towards this goal (e.g. used by disease prediction algorithms?)
-
