Algorithms.

The algorithm for computing trajectories is based on the method for analyzing trajectories proposed by Jensen et al [1]. From a high level perspective, it consists of three steps:

1. Select medical event pairs A → B with a relative risk score (RR) > 1.0:
   • ∀ possible disease pairs A → B calculate the RR, i.e. the chance for developing B when diagnosed with A.
   • The RR is calculated via a sampling experiment comparing p-values to guarantee a fair estimation of the RR based on our data.
   • If both A → B > 1.0 and B → A > 1.0, perform a directionality test with a binomial experiment to select one pair.
   • Only keep the pairs A → B for which there is a minimum number of patients.
2. Build trajectories A → B → … → Z by chaining pairs:
   • E.g. A → B and B → C form a new trajectory A → B → C.
   • For each extension A → B → C check that there is a minimum number of patients.
3. Cluster trajectories for visual inspection, pattern detection, and hypothesis generation.

The details of these steps are discussed next.

1. Selection of medical event pairs.

The selection of medical event pairs requires calculating the relative risk scores for all possible pairs, i.e. calculating for each event pair A → B what the probability is that a medical event B follows a medical event A [1, 15]. For example, if the RR is 6, it means that B occurs 6x more for patients who have previously been diagnosed with A than those who have not been diagnosed with A.

The usual way to calculate relative risk scores is to set up a prospective cohort study comparing patients exposed to A with patients not exposed to A and then counting the occurrence of event B in both patient groups to calculate the probability for A → B. However, the idea here is to calculate the relative risk scores with the data available in the patient histories. The strategy is to setup random sampling experiments on this data to have a fair simulation of cohort studies.

Simplified pseudo code for calculating relative risk scores for all possible medical event pairs is shown in Listing 3.

∀ pair A→B:

 (P1 ← get patients exposed to A

 P2 ← sample #P1 similar patients ¬exposed to A

 Count #(A→B) in P1, Count #(¬A→B) in P2

 Calculate RR)

 × 10.000

Calculate p-value

Listing 3. Pseudo code relative risk score (RR) calculation.

For all possible event pairs (e.g. all combinations of ICD diagnosis codes), we calculate the number of patients exposed to a first medical event (A). Subsequently, we sample similar patients in terms of age, sex, localization, etc that are not exposed to that medical event (¬A). It is important to sample similar patients to avoid Simpson’s paradox [1]. For both patient groups, we count if a second event (B) follows the first medical event or absence thereof. Using these counts, we can calculate the relative risk score (RR). We repeat this experiment a number of times to calculate p-values and guarantee a fair random selection of patient groups to simulate the cohort study.

The repeated random sampling process to calculate RR scores is computationally expensive. The sampling loop needs to be processed for each possible combination of medical events. If we assume for example that the medical events are diagnoses encoded in ICD10, this means attempting a combination of 1064 × 1064 diagnosis codes for level 3 in de ICD10 hierarchy. The computation also relies on a random number generator to sample the patient groups (P2), which is computationally a costly operation.

We therefore parallelize the calculation of RR scores in the PTRA implementation. The RR computation for each pair A → B can occur independently. We can therefore split up the iteration space of our loop that calculates RRs and pass these iteration groups to independent parallel tasks. Profiling reveals that random number generation is computationally the most costly step in the RR calculation. We therefore use a custom library for random number generation rather than using the one from the Go standard library [16].

We refer to the source code [17] for more details on the RR calculation we left out in our discussion so far. For example, there is a statistical step for filtering A → B pairs upfront so they are not checked in the sampling experiment. For input data that is in ICD format, there are filters to remove entire ICD categories from the analysis (e.g. related to pregnancy). Also, counting A → B pairs must take into account timing constraints between diagnoses (e.g. maximum 5 years, minimum 6 months). For ICD codes, the ICD hierarchy must be taken into account as well as the level the analysis is run with.

2. Building trajectories.

The first step for building trajectories involves selecting medical event pairs based on the value of the RR score for that pair. The selected pairs are the starting points for chaining combinations of pairs into trajectories. By default, all pairs with a RR score >1.0 are selected for trajectory building. Additionally, if a pair A → B and its reverse B → A both have a RR score >1.0, a binomial experiment is used to test the directionality and select either the pair or its reverse for trajectory building. Finally, for a pair A → B to be considered for trajectory building, there also has to be a minimum number of patients in the population that are diagnosed with A → B [1].

To parallelize the trajectory building, the selected diagnosis pairs are divided among a number of parallel processes. These pairs are seen as partial trajectories. Each process subsequently attempts to extend its partial trajectories by trying to combine them with all of the selected pairs. For example, a partial trajectory A → B can be combined with B → C because they overlap with diagnosis B, provided additional constraints are satisfied. These constraints are 1) that the extended trajectory contains a minimum number of patients that follow it and 2) that the time between A → B and B → C satisfies a certain time span (e.g. between 6 months and 5 years). The extended trajectories found this way are subsequently checked for extension until no more extensions are possible. If extension is not possible, a partial trajectory is finalized and returned as a result.

Each process uses a thread-local stack for storing its intermediate trajectories. The finalized trajectories are also written to a thread-local buffer. After all processes are finished, local results are merged. The parallelism we describe here can be implemented with a fork/join pattern. Care must be taken that partial trajectories are copied when extended rather than destructively written since they may be extended multiple times. We found our fork/join-based parallelization to run efficiently on a single shared memory multicore server, whereas the Jensen method needs cluster computing for processing similarly sized data sets [1].

3. Clustering trajectories.

The Jensen algorithm for computing trajectories produces trajectories with a lot of overlap [1]. Therefore clustering of the trajectories is used to group together similar trajectories for visual inspection. The clusters are visualized by representing all trajectories in the clusters as a graph: diagnoses are the nodes of the graph and trajectory transitions are the edges. This type of visualization of trajectory clusters can help interpreting the trajectories and generating medical hypotheses [1–5].

The original algorithm by Jensen et al. clusters trajectories as shown in the pseudo code in Listing 4.

#Cluster the event codes

G ← make a graph: event codes (D) = nodes, RR pairs = edges

#Place a similarity score on each edge ∈ G using the Jaccard index:

∀ Edge A → B ∈ G:

 Jaccard(A → B)←(#T|A → B ∈ T)/(#T|A ∈ T + #T|B ∈ T − #T|A → B ∈ T)

#Cluster G with MCL, assigns each T to a cluster C_i

C₁, C₂, …C_N ← MCL(G)

# Assign each trajectory T to a cluster C_i

T ∈ C_i if ∀D ∈ T|D ∈ C_i

Listing 4. Pseudo code clustering strategy in Jensen et al [1].

The strategy is to first cluster the event codes. A graph is created that places each event code onto a unique node. The selected pairs (from step 2) are used to create edges between event nodes. For each edge in the graph, a similarity score is calculated using the Jaccard index. As the pseudo code shows, the Jaccard index is calculated for a pair A → B by checking how many trajectories T exist that contain that pair and compare it to how many trajectories exist that contain its constituents A and B, but not the pair itself [1].

Once each edge is assigned a Jaccard index, the medical event graph is clustered using the MCL [18] algorithm. Finally, trajectories are assigned to their clusters. A trajectory is assigned to a cluster if all of the event codes used in the trajectory are a part of the cluster. In practice, this means certain trajectories are not assigned to any cluster, as event codes are unique in the graph [1]. The solution proposed by Jensen et al [1] is to manually merge clusters and assign trajectories that are otherwise not assigned to the merged clusters. We found that this method becomes impractical for analyzing specific data sets. For example, as we discuss in later sections, the strategy leaves up to 2/3 of the trajectories in TriNetX Dataworks to be assigned manually.

We propose a different clustering strategy than the one described by Jensen et al. Pseudo code is shown below in Listing 5.

# Assign each trajectory T a unique ID

G ← make a graph: trajectory IDs = nodes, fully connected

# Place a similarity score on each edge ∈ G

# using the Jaccard index to directly compare T

# View T as sets of medical event codes

∀ Edge T₁ → T₂ ∈ G:

 Jaccard(T₁, T₂ ← |T₁ ∩ T₂|/(|T₁| + |T₂| − |T₁ ∩ T₂|)

# Cluster G with MCL, assigns each T to a cluster C_i

C₁, C₂, …C_N ← MCL(G)

# Assign T to cluster C if TID ∈ C

T ∈ C_i if TID(T)∈C_i

Listing 5. Pseudo code PTRA clustering strategy.

The idea behind our clustering strategy is to directly cluster the trajectories. To do this, we assign each trajectory (T) a unique trajectory ID (TID). We make a graph that has as nodes all of the trajectory IDs. The graph is fully connected so that we can compare all trajectories to all trajectories. We place a similarity score on the edges of this graph by computing the Jaccard index of the trajectories the edge connects. To compute the Jaccard index of two trajectories, we think of them as sets of event codes. For an edge T₁ → T₂, the Jaccard index is then the ratio of the number of event codes in the intersection of T₁ and T₂ versus the number of event codes unique to T₁ and T₂. This is the graph we cluster with MCL (or another clustering algorithm). The cluster of a trajectory is then the cluster to which its trajectory ID (i.e. graph node) is assigned. The advantage of this clustering strategy is that all trajectories are guaranteed to be assigned to a single cluster.

TriNetX Dataworks data set for bladder cancer patients.

The data used in this study was collected in December 2021 from the TriNetX Dataworks Network [19], which provided access to electronic medical records (diagnoses, procedures, medications, laboratory values, genomic information) from approximately 110,627,025 patients from 74 healthcare organizations. TriNetX, LLC is compliant with the Health Insurance Portability and Accountability Act (HIPAA), the US federal law which protects the privacy and security of healthcare data, and any additional data privacy regulations applicable to the contributing HCO. TriNetX is certified to the ISO 27001:2013 standard and maintains an Information Security Management System (ISMS) to ensure the protection of the healthcare data it has access to and to meet the requirements of the HIPAA Security Rule. Any data displayed on the TriNetX Platform in aggregate form, or any patient level data provided in a data set generated by the TriNetX Platform, only contains de-identified data as per the de-identification standard defined in Section §164.514(a) of the HIPAA Privacy Rule. The process by which the data is de-identified is attested to through a formal determination by a qualified expert as defined in Section §164.514(b)(1) of the HIPAA Privacy Rule. Because this study used only de-identified patient records and did not involve the collection, use, or transmittal of individually identifiable data, this study was exempted from Institutional Review Board approval.

We extracted a cohort of bladder cancer patients from TriNetX Dataworks [19]. Bladder cancer is one of the main driver cases for Project ATHENA (Augmenting Therapeutic Effectiveness through Novel Analytics) [20] we participate in. Bladder cancer has multiple stages and bladder cancer patients have a high risk of disease recurrence and progression to different stages. Hence studying bladder cancer trajectories is useful. To the best of our knowledge, bladder cancer trajectories extracted from medical event histories have not yet been studied in related work.

We identify bladder cancer patients as patients who have a record of a diagnosis “Malignant neoplasm of bladder” (ICD10: C67 or ICD9: 188) or a diagnosis “Personal history of malignant neoplasm of bladder” (ICD10: Z85.51 or ICD9: V10.52). We extracted this bladder cancer cohort in December 2021 and this gave us a data set with around 128 thousand patients and around 47 million timestamped ICD9/ICD10 diagnosis codes for those patients. As a point of comparison, for example, the data set used by Jensen et al [1] has around 6.2 million patients and 101 million diagnosis codes. Hence our data set has fewer patients, but far denser diagnosis histories. In addition to ICD codes, the TriNetX Dataworks data set also contains data for bladder cancer-specific treatments such as MVAC chemotherapy, radical cystectomy, and intravesical therapy. We assign these treatments unique medical event codes so they can be used in addition to the ICD codes as inputs to PTRA.

The PTRA tool.

While PTRA is designed as a library for implementing analyses of medical event histories, it also implements a concrete use case. A command-line interface (CLI) is implemented to use PTRA as a tool for analyzing data extracted from the TriNetX Dataworks database [19]. It is useful to go over this CLI because it can help to understand the experiments we set up to evaluate PTRA.

The PTRA CLI for analyzing TriNetX is shown below in Listing 6:

ptra patientInfoFile diagnosisInfoFile diagnosesFile outputPath

 —nofAgeGroups nr —minPatients nr

 —maxYears nr —minYears nr

 —maxTrajectoryLength nr —minTrajectoryLength nr

 —iter nr —lvl nr

 —pfilters [age70+|age70-|male|female|NMIBC|MIBC|mUC|…]

 —tfilters neoplasm|bc

Listing 6. PTRA CLI tool for analyzing TriNetX data.

The CLI lists 4 required parameters: 1) a CSV input file with patient information such as sex, age, etc; 2) a diagnosis info file that maps diagnosis codes onto medical descriptors: e.g. XML file with ICD10 hierarchy from the WHO [10] or a CSV with CCSR categorization of the ICD10 hierarchy [11]; 3) a CSV file with timestamped ICD10/ICD9 diagnoses for the patients; 4) a path where to write the output.

There are optional parameters to further steer the trajectory building. The patient sampling for calculating RR scores compares patients similar in terms of age and sex. The age range for two patients to be considered similarly aged can be configured (--nofAgeGroups). The minimum number of patients in a trajectory can be set. The maximum and minimum time between diagnoses can be set, as can the maximum and minimum length of a trajectory. The number of iterations for the sampling process can be set (--iter). The level of the ICD diagnosis codes within the ICD hierarchy can be set so that ICD codes may be combined for analysis. Finally, there are filter options. A first set of filters (--pfilters) can be used to restrict the patients for which trajectories are calculated: e.g. only people above 70 years old, only females, only patients who are in the non muscle-invasive bladder cancer (NMIBC) stage, etc. Then there is a second filter option (--tfilters) that can be used to reduce the trajectory output, e.g. to include only trajectories with a cancer-related output.

There are additional options that are not listed in listing 6. We refer to the README with comprehensive documentation on our github repository for more details [17]. The PTRA tool is distributed as a single executable for Linux and can be compiled from source or downloaded directly from the repository.

A PTRA run creates multiple output files:

1. a .tab file with the selected diagnosis pairs and their relative risk scores.
2. a .tab file with the trajectories found.
3. a folder with clustered trajectory output. This folder contains up to four files:
   1. (a) a .csv file with cluster information linking patient, trajectory, and cluster.
   2. (b) a .csv file linking patient to TriNetX patient identifier and selected or derived TriNetX information about that patient (sex and age at event of interest). The two .csv files can be used to plot statistics per cluster (Julia script included in PTRA source code).
   3. (c) two graph modeling language (.gml) files with clustered trajectories encoded as a graph per cluster. The .gml format is a standardized format for storing graphs and such files can be visualized with external tools such as yED. There is one .gml file where trajectory transitions are annotated with the number of patients in the trajectory so far, and a second .gml file where the trajectory transitions are annotated with the relative risk score for the diagnosis pairs. An example is shown in Fig 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Example trajectory cluster extracted from TriNetX Dataworks (bladder cancer population).

The graph on the left is annotated with the number of patients for consecutive diagnoses in a trajectory. The graph on the right represents the same trajectory cluster, but it is annotated with the RRs for each diagnosis pair instead. The graph on the left has more edges than the graph on the right because it shows all distinct trajectories annotated with their patient numbers. For example, different trajectories end up in “Other chronic obstructive pulmonary disorder” while having different starting points: “Nicotine dependence” or “Other respiratory disorders”.

https://doi.org/10.1371/journal.pdig.0000384.g001

For more details on the outputs generated by PTRA, we refer to our README on github [17].

PTRA cluster interpretation.

An example output for clustering trajectories with PTRA is shown in Fig 1. It shows the same cluster twice: once annotated with the number of patients and once annotated with the relative risk scores between diagnoses. The thickness of the connecting lines increases with the number of patients or relative risk scores respectively. The cluster shows that there are multiple trajectories leading to specific diagnoses. For example, there is a trajectory from “Nicotine dependence” to “Emphysema” to “Other chronic obstructive pulmonary disease”, but also a trajectory from “Gastro-esophageal reflux disease” to “Asthma” to “Other chronic obstructive pulmonary disease”. The first trajectory shows the negative impact of smoking, whereas the second trajectory reveals that asthma is often first misdiagnosed as gastro-esophageal reflux disease.

In order to interpret a PTRA cluster, there are a number of steps we recommend:

1. Identify the key diagnoses in a cluster. These are the diagnoses with the most connections in the graphical output. For example, in Fig 1 these are “Other chronic obstructive pulmonary disease” and “Emphysema”.
2. Identify the common trajectories by following the arrows that lead up to or start from key diagnoses.
3. Identify the diagnoses with the highest risk to predate or follow a (key) diagnosis. These can be spotted by looking at the thickest lines in the graphical output. For example, in Fig 1, the highest risk score is found for the transition of “Emphysema” to “Other chronic obstructive pulmonary disease”. It shows that patients diagnosed with “Emphysema” have three times more chance to be afterwards diagnosed with “Other chronic obstructive pulmonary disease” than patients that have not been diagnosed with “Emphysema”.
4. Check the number of patients from one diagnosis to another and compare it to the relative risk score. For example, in Fig 1, we see the highest number of patients between “Gastro-esophageal reflux disease” and “Asthma”. In comparison, the relative risk score is not as high as for other diagnoses. This suggests there are many patients diagnosed with “Asthma” that are not diagnosed with “Gastro-esophageal reflux disease”. In this case, this hints that the latter may be a misdiagnosis for the former.

The PTRA framework.

The PTRA package is organized as a framework that defines APIs to implement trajectory analyses for medical event histories. We discuss code organization, how to add data filters, and the protocols for implementing a new use case in S1 Appendix.

Experiments.

We set up different runs with PTRA to analyze the bladder cancer cohort we extracted from TriNetX Dataworks. As a point of reference, we set up a run that extracts all trajectories for the full cohort. This is easiest to compare to the experiments from the papers on the original methods which focus on processing complete diagnosis histories of all patients in a population [1, 2, 8].

We additionally set up an experiment that extracts bladder cancer-specific trajectories. For this experiment, we use PTRA-specific features. Concretely, we restrict the output to return trajectories that contain at least one bladder cancer-related diagnosis. This comprises ICD10 codes C67, C77, C78, C79, as well as medical procedures such as radical cystectomy, MVAC chemotherapy, and intravesical therapy that are registered separately in TriNetX Dataworks. The PTRA algorithm automatically maps these procedures onto unique medical event codes for this purpose. Additionally, we set up PTRA to remap the ICD codes onto CCSR categories for assisting medical interpretation of the trajectories.

Table 1 gives an overview of the parameter settings for the different runs. For overlapping parameters, we chose similar values as those used in related work [1–6, 8].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Experiment parameters for runs generating 1) all trajectories based on full ICD histories and 2) bladder cancer-specific trajectories found using ICD histories and bladder cancer-specific treatments.

https://doi.org/10.1371/journal.pdig.0000384.t001

Server and software versions.

All of our experiments ran on a 36-core server, consisting of two 18-core Intel Xeon E5–2699v3 Haswell CPUs clocked at 2.3Ghz. The server is equipped with 256GB RAM and two 400GB SSD disks for storing local data. The machine runs Ubuntu 20.04.4 LTS Linux. PTRA was compiled using Go 1.18.4. For clustering, we used the MCL program version 11–294. The trajectory output of PTRA (in .gml format) was visualized using yEd 3.21.1. Plotting and additional analysis was done with Julia 1.7.3.

General disease development patterns in bladder cancer patients in TriNetX Dataworks.

In our first experiment, we generate trajectory clusters for the full cohort of bladder cancer (BC) patients. This means that we calculate trajectories using the full diagnosis histories of all patients in the cohort and collect all trajectories found this way. This is similar to the approach taken in related work [1, 2, 8].

This approach reveals clusters of highly prevalent, well-known chronic diseases such as cardiovascular disease, diabetes, dorsalgia, etc, which co-occur in an aging population such as BC patients. These clusters are not necessarily related to BC and this makes it difficult to zoom in on BC-specific development patterns. Nonetheless, we can assign medically valid interpretations to these clusters, showing the PTRA algorithm produces sensible results.

For the full BC cohort, PTRA finds 17 clusters (S2 Appendix). The 5 largest clusters found, have the following central diagnoses: heart disease (cluster 0), soft tissue disorder (cluster 1), musculoskeletal disorder (cluster 3), cardiorenal disease (cluster 4), and anxiety/sleep disorder (cluster 5). When examining age and sex-related characteristics per cluster, we see they confirm some well-known disease characteristics:

• the percentage of males is higher in the heart failure cluster.
• the percentage of females is higher in the clusters that involve anxiety disorder.
• the mean age of diagnosis of bladder cancer is similar across all clusters: on average 70 years of age.

While the clusters are clinically correct, the specific findings do not represent new knowledge per se, as the most general clusters of diseases tend to dominate in the output. A lot of clusters are also dominated by overarching and unspecific ICD codes such as “Other and unspecified soft tissue disorder, not elsewhere classified” and “Other joint disorders”. Additionally, no cluster includes BC as a diagnosis event. This is logical, as each patient is diagnosed with BC in the input data. Therefore the algorithm cannot sample comparison groups without BC (cf. Listing 3) and the statistics cannot pick it up as an event of interest. This makes it difficult to study BC-specific development patterns.

In our second experiment, we show how PTRA-specific features make it easier to study BC-specific trajectories. Specifically:

• We make use of CCSR categorization of ICD10 codes. The Clinical Classifications Software Refined (CCSR) aggregates ICD10 codes into clinically meaningful categories. This bypasses the problem of clusters being dominated by overarching or unspecific diagnoses.
• We include BC-specific treatments to the diagnosis histories for trajectory building. This gives a handle on identifying BC-specific trajectories. We also changed the minimum time between diagnoses from six months to one day to capture the fact that BC treatments may follow each other quickly.
• We use the trajectory filter mechanism to restrict trajectory output to trajectories that include at least one known bladder cancer-related diagnosis (ICD code) or treatment. This steers the PTRA algorithm to cluster and output only trajectories related to BC.

Bladder cancer development patterns in TriNetX Dataworks.

Approximately 75% of patients with BC present in the early stage (non-muscle-invasive bladder cancer—NMIBC) of the disease, meaning that the disease is confined to the mucosa (stage Ta, carcinoma in situ [CIS]) or submucosa (stage T1). In this group, recurrence and progression remain problematic. Despite intensive treatment and follow-up schedule, these patients have an elevated risk for disease recurrence and a moderate to elevated risk for disease progression of up to 35–55% at a year follow-up [21]. Management of NMIBC consists of TURBT (operative removal of tumor) and bladder instillations with BCG (Bacillus Calmette-Guérin) plus intensive follow-up and maintenance BCG therapy. The rate of BCG non-responders amounts to 40%. This explains why the European Association of Urology states in their most recent guidelines that radical cystectomy should be offered to the very high-risk patients, even in early stages of the disease [22]. Removal of the bladder drastically diminishes the patients? quality of life, making it incredibly important to optimally discriminate between high- and very high-risk patients. Additionally, beforehand determination of response to treatment with intravesical BCG instillations in BC would increase the survival of patients while eliminating time loss and unnecessary use of medications. Management of late-stage (muscle-invasive bladder cancer—MIBC) disease consists of radical cystectomy with or without chemotherapy, depending on eligibility of the patient for surgery/chemotherapy.

We need to be able to better stratify BC patients. In this perspective, there is a need to:

1. Identify BC patients who are at the highest risk for disease progression.
2. Predict treatment response.

In pursuit of these goals, we analyze the PTRA trajectories found for the BC patients in the TriNetX Dataworks database. We use the PTRA trajectory filters to only output and cluster trajectories that contain a BC-related diagnosis or treatment.

Using PTRA, 10 clusters of disease trajectories were discovered. The dominant diagnostic event in each of the clusters is secondary malignancy (cluster 0, 3 [respiratory cancer] and 5) (Fig 2), radical cystectomy (cluster 1, 8 and 9) (Fig 3), and MVAC (Methotrexate-Vinblastine-Doxorubicin [Adriamycin]-Cisplatin) chemotherapy (cluster 2, 4, 6 and 7) (Fig 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Clusters for secondary malignancy.

PTRA clusters with secondary malignancy as central diagnosis (in bold). The clusters show different trajectories leading up to secondary malignancy. For example, cluster 0 shows trajectories where hematological and drug-induced disturbances precede secondary malignancy, whereas cluster 5 shows infection-related events. The clusters suggest a different development of events after the secondary malignancy diagnosis: cerebrovascular disease, epilepsy, and paralysis for cluster 0, and respiratory insufficiency for cluster 5.

https://doi.org/10.1371/journal.pdig.0000384.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Clusters for radical cystectomy.

PTRA clusters showing potential complications after radical cystectomy (in bold). Ranked from highest to lowest risk, this includes: post-procedural or postoperative digestive system complication, reflux disease, peritonitis and intra-abdominal abscess, postoperative skin complication, urinary tract infection and gastrointestinal and biliary perforation. Also rare, but severe complications such as septicemia and pulmonary embolism are identified, albeit with lower relative risk scores than the other complications. These findings seem to correlate with complications of radical cystectomy described in literature.

https://doi.org/10.1371/journal.pdig.0000384.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Clusters for MVAC chemotherapy and effects.

PTRA clusters with MVAC chemotherapy as a central event (in bold). The clusters show follow-up events that are known side-effects described in literature such as: drug-induced toxicity, acute phlebitis, thromobphlebitis, and thromboembolism, fungal infections, and polyneuropathy. However, the clusters also highlight a high risk of rheumatoid arthritis with a relative risk score of 14.15, which is a remarkable finding as flare up of auto-immune diseases with MVAC chemotherapy is not assumed to be conventional.

https://doi.org/10.1371/journal.pdig.0000384.g004

Cluster 0, 3 and 5 show events preceding the diagnosis of secondary malignancy, presenting risk factors or biomarkers for detecting secondary malignancy (Fig 2). Multiple types of hematological deviations are identified as preceding events in cluster 0: aplastic anemia having the highest risk ratio (RR) (2,3), followed by fluid/electrolyte disturbances (2,26), diseases of white blood cells (2,07) and nutritional anemia (1,48). This cluster also revealed drug-induced or toxicity-related conditions preceding secondary malignancy with a relatively high RR (2,59). This phenomenon seems logical because the need for terminating treatment due to side effects diminishes the chances of treatment success, which in turn could lead to disease progression. Cancer of the small intestine is the event preceding secondary malignancy with the highest RR (3,19). This finding is surprising because the small intestine is not the primary location for metastasis and primary cancer of small intestine itself is a very rare event. Most often BC will spread to lymph nodes, bones, lung, liver, and peritoneum [23, 24]. Due to the nature of the dataset that is used (TriNetX), we lack granularity in the terminology to clearly explain the underlying disease physiology of these findings. We require more detailed information on the definitions of used terms. However, the dataset used to develop this algorithm does not contain this detailed amount of information. More research exploring this algorithm on different datasets is needed.

When comparing cluster 0 and cluster 5, both having secondary malignancy as central diagnosis, different types of trajectories are revealed. Within cluster 0, multiple preceding events are hematological and drug-induced disturbances while cluster 5 shows more infection-related events (pericarditis, intestinal infection, peritonitis and intra-abdominal abscess etc.). This seems to have an influence on events occurring after diagnosis of secondary malignancy; cluster 0 leading to cerebrovascular disease, epilepsy and paralysis and cluster 5 leading to respiratory insufficiency.

Cluster 1, 8 and 9 reveal potential complications of radical cystectomy (Fig 3). When ranked from most to least likely: post-procedural or postoperative digestive system complication (RR 1,95), reflux disease (RR 1,78), peritonitis and intra-abdominal abscess (RR 1,75), post-procedural or postoperative skin complication (RR 1,27), urinary tract infection (RR 1,2) and gastrointestinal and biliary perforation (RR 1,13). Severe complications such as septicemia and pulmonary embolism are also identified, though rare compared to the other events (RR 1,02 and 1,08). These seem to correlate with complications described in literature.

Cluster 2, 4 and 6 reveal MVAC chemotherapy and its effects (Fig 4). The clusters show that rheumatoid arthritis (RA) and related disease is one of the most common events after MVAC therapy (with RR of 14,15), along with diseases of the mouth (RR 2,44), drug-induced toxicity (RR 2,54), acute phlebitis, thrombophlebitis, and thromboembolism (RR 1,86) and others such as, fungal infections (RR 1.36), polyneuropathy (RR 1,78) etc. Most of these events are known side-effects described in literature [25]. However, the high risk of RA after MVAC is noteworthy. Cancer and autoimmune disease are closely related, and many therapeutic antibodies are widely used in clinics for the treatment of both diseases. At the same time, certain immunotherapies used in cancer treatment can induce autoimmune reaction [26]. With MVAC chemotherapy, flare up of auto-immune diseases is not conventional. Specifically, because Methotrexate is used in low doses for treating RA. A clinical correlation between the RA and MVAC does exist, the order of events in our trajectories, however, is remarkable.

Using PTRA, we can also analyze the patient characteristics of each of our clusters (Fig 5). Gender analysis shows that in clusters with MVAC as dominant diagnostic event, the percentage of females is higher, while clusters with radical cystectomy are male dominated. This is consistent with earlier studies, where female gender is reported as a risk factor for disease recurrence, progression, and cancer-specific mortality requiring more advanced treatments [27]. The literature suggests that female gender in patients diagnosed with BC is associated –though inconsistently– with higher stage at presentation and poorer outcomes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Gender distribution per cluster.

This graph shows for each PTRA cluster the percentage of female patients versus the percentage of male patients that fall into that cluster. It shows, for example, for the MVAC clusters (clusters 2,4,6 from Fig 4), that the percentage of females is higher than the percentage of males. In contrast, the clusters with radical cystectomy as central event (clusters 1, 8, 9 from Fig 3), contain more males than females. This is consistent with literature that reports female gender as a risk factor for disease recurrence, progression, and cancer-specific mortality requiring more advanced treatments.

https://doi.org/10.1371/journal.pdig.0000384.g005

Age analysis in Fig 6 reveals that in clusters were MVAC chemotherapy is a dominant diagnosis event, the average age of onset of disease is lower compared to other clusters (+/-63 years vs +/-70 years). In contrast, older adults have a higher mortality due to BC than younger individuals. We would therefore expect our older population of BC patients to have progressive disease and need additional treatment with surgery or chemotherapy. However, toxicity is a major concern with MVAC chemotherapy, particularly in an elderly population with multiple comorbidities. Therefore, an older population is more often ineligible for MVAC therapy. This explains why the mean age in these clusters is lower.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Average age of onset of BC per cluster.

This graph shows fro each PTRA cluster the average age a patient in that cluster is diagnoses with bladder cancer. It shows for the MVAC clusters (clusters 2, 4, 6 from Fig 4) that the average age of onset of disease is lower compared to other clusters: +/-63 years vs +/-70 years. We assume this is because toxicity is a major concern with MVAC chemotherapy, especially in an elderly population with multiple comorbidities. Therefore older patients are often ineligible for MVCA therapy.

https://doi.org/10.1371/journal.pdig.0000384.g006

Limitations.

While working with ICD-10 codes can provide valuable information about patient trajectories, there are several limitations we need to address:

• Coding errors and inconsistencies: Assigning ICD-10 codes can be a complex task, and errors or inconsistencies may occur during the coding process. These errors can lead to inaccuracies in patient trajectory analysis, potentially impacting the interpretation of data. An illustrative example is the diagnosis of sepsis. Sepsis, a life-threatening blood infection, is typically diagnosed based on blood cultures. Given the urgent need for immediate intervention in sepsis cases, the coding aspect is sometimes overlooked and not added to the electronic medical records retrospectively [28]. Another observation from our experiments highlighted the inconsistencies of disease progression diagnoses. We noticed instances where patients were receiving therapy for advanced-stage cancer, while their records still indicated an early-stage cancer code. The presence of coding inconsistencies, as exemplified by this discrepancy, can result in substantial information loss, underscoring the importance of addressing this limitation.
• Lack of granularity: ICD-10 codes provide a standardized system for classifying diseases and health conditions, but they often lack the level of detail necessary to capture the complexity of patient trajectories. The codes may not fully capture the nuances of a patient’s specific condition, variations in disease severity, or the progression of symptoms over time. In our experiment, we encountered this limitation when it came to diagnosing secondary malignancy. This diagnosis can be interpreted in different ways, potentially indicating the development of a new primary cancer or it can indicate the progression of an existing disease with the occurrence of metastases. Due to the lack of more detailed information, we were unable to draw any definitive conclusions.

To overcome the limitations associated with ICD-10 codes and the inconsistencies we encountered, it is crucial to augment the analysis with additional data sources and clinical information. This approach will ensure a more comprehensive understanding of patients’ healthcare journeys. In future work, the following steps can be taken to address specific inconsistencies:

• Sepsis case: Incorporate supplementary data such as lab results and the diagnosis of bacteria in blood cultures to aid in assigning the sepsis diagnosis. Additionally, consider including information on antibiotics administered as treatment for sepsis. By integrating these data points, a more accurate and comprehensive picture of the sepsis diagnosis can be obtained.
• Metastasis case: Implement an automated system that adjusts the disease stage (early/late stage cancer) once a treatment for late-stage cancer is recorded. This would ensure that the coding accurately reflects the patient’s current disease status and progression. By incorporating these strategies and leveraging additional data sources, researchers and healthcare professionals can mitigate the limitations associated with ICD-10 codes and gain a more holistic understanding of patients’ healthcare journeys.