Data Harmonization for Collaborative Research Among Australian and US Registries: A Case Study in Medullary Thyroid Cancer (MTC)

ABSTRACT

Background

Medullary thyroid cancer (MTC) is a neuroendocrine tumor comprising approximately 1%–2% of all thyroid malignancies. The rarity and more aggressive biology of MTC requires robust sample sizes to enhance our understanding of this complex disease. Harmonization is the process of standardizing raw data from multiple sources, by resolving differences in format and terminology, to create a unified dataset that can be analyzed for a common purpose. The aim of this project was to assess the feasibility of collaboration, data mapping, and harmonization among clinical sites investigating MTC internationally.

Methods

The Maelstrom guidelines were used to perform retrospective data harmonization from three clinical networks in Australia and the Unites States for adult patients with MTC, between 2018 and 2021. Data received were categorized as an exact, close, or low match. Exact and close matches were combined to form a harmonized dataset. A logistic regression analysis was then performed to determine pre‐operative factors associated with the presence of cervical lymph node metastases.

Results

Data were received from three separate clinical networks. This comprised 114 patients, 17 hospitals and 4674 data points. The completeness of data received ranged from 57.4% to 97.3%. Overall, 80.8% of data received were suitable for harmonization including basic demographics, basis of diagnosis, genetic testing (but not results), select clinical findings, pre‐operative investigations, operative details, histopathology, and TNM staging. The prevalence of palpable lymph node involvement at presentation in the harmonized dataset was 15.8%. Younger patients (less than 55 years) and patients with abnormal nodes on ultrasound were strongly associated with cervical lymph node metastases. Conversely, patients with an incidental diagnosis of MTC had markedly lower odds of presenting with cervical lymph node metastases.

Conclusion

Data mapping and harmonization across national and international sites is feasible and enables meaningful modeling that would not be possible with individual datasets. The Maelstrom guidelines provide a useful template regarding how to achieve this efficiently. This manuscript is a white paper for clinicians and researchers studying rare diseases, such as MTC, regarding how to share heterogeneous raw data and collaborate with other clinical sites.

Medullary thyroid cancer (MTC) is a neuroendocrine tumor comprising approximately 1%–2% of all thyroid malignancies. The rarity and more aggressive biology of MTC requires robust sample sizes to enhance our understanding of this complex disease. Harmonization is the process of standardizing raw data from multiple sources, by resolving differences in format and terminology, to create a unified dataset that can be analyzed for a common purpose. The aim of this project was to assess the feasibility of collaboration, data mapping, and harmonization among clinical sites investigating MTC internationally.

1. Introduction

Medullary thyroid cancer (MTC) is a rare neuroendocrine tumor of the parafollicular C‐cells within the thyroid gland [1, 2]. The C‐cells are located throughout the thyroid and are most numerous in the upper two‐thirds of the gland [3]. Most episodes of MTC are sporadic; however, 20% may be predicted by autosomal dominant heritance of a Rearranged During Transfection (RET) mutation [4]. Although MTC comprises only 1%–2% of all thyroid malignancies [5], it has a high mortality burden (up to 15%) [6, 7, 8].

Patients with MTC may present with a palpable mass for screening based on the presence of a hereditary mutation or are discovered incidentally via post‐operative histopathology. The clinical course may be chronic and can be highly variable. For clinicians and researchers, this necessitates long‐term follow‐up and robust sample sizes in order to investigate clinically relevant aspects of the disease.

Clinical quality registries are invaluable in this field and there are multiple existing clinical registries and institution‐specific databases for MTC. However, even nationwide registries studying rare diseases remain limited due to small annual numbers of eligible patients and smaller clinical events. The ability to consolidate raw data from multiple sources, nationally and internationally, would potentially enable more meaningful and higher powered clinical research to be conducted.

Data mapping is the systematic process of establishing semantically equivalent relationships between data elements in a source dataset. The aim of data mapping is to assess whether data from different sources is comparable. Data harmonization is the process of integrating, standardizing, and aligning existing raw data from multiple sources. The aim of harmonization is to combine data seamlessly and efficiently without data loss.

Data mapping and harmonization are used in a range of industries to consolidate data processes and achieve equivalence. In healthcare, it enables sharing of clinical information between health professionals and to create larger individual participant data (IPD) mega‐datasets. Harmonized datasets are preferable to aggregate datasets, in terms of quality (reduced risk of bias, greater statistical power, better handling of heterogeneity, option for subgroup analyses) and applicability. However, they are costly, labor intensive, and raw data are not always accessible. Systematic guidelines for retrospective data mapping and harmonization have been developed by Maelstrom Research and its partners [9].

This study had two aims. The first was to assess the feasibility of data mapping and harmonization across three high volume tertiary referral centers for patients with MTC. Secondly, as an example of the utility of harmonization, to determine the risk factors for cervical lymph node metastases (via logistic regression analysis of our harmonized dataset). Our hypothesis was that academic collaboration is feasible and can lead to more meaningful information, although challenges such as data incompleteness, privacy, and inefficiency need to be considered.

2. Methods

2.1. Clinical Networks

Three clinical networks, comprising 17 sites, were identified and invited to participate in this study. The Monash University Endocrine Surgery Unit (MUESU) is a collaborative clinical outcomes database of endocrine surgery patients in association with all Monash University hospitals. The Sydney Endocrine Surgery Unit (SESU) is also a clinical outcomes and biologic database of endocrine surgery patients in association with the University of Sydney and Kolling Institute Tumor Bank. Cleveland Clinic's Department of Endocrine Surgery is a high‐volume endocrine surgery unit in association with the tertiary academic center's Integrated Surgical Institute. Its database is a hybrid data repository for endocrine surgery patients, which links to the entire enterprise's electronic medical records system (EPIC) [10].

2.2. Maelstrom Guidelines

The Maelstrom guidelines [5] are a comprehensive methodological protocol for data harmonization. They were developed by Maelstrom research and partners following an iterative process using surveys, workshops, and pilot programmes. The guidelines comprise six discrete steps and a 14‐point process checklist. The first three steps are to define the research question and objectives, assemble pre‐existing knowledge, and to determine the data items which are feasible to be harmonized. The final three steps are to process the data and generate a harmonized dataset, estimate quality, and disseminate the results. The guidelines present a balanced perspective for harmonization by also including desirable adjunct components (e.g. collaborative framework, expert opinion, valid data input and output methods, rigorous documentation, respect for all stakeholders) and potential barriers.

2.3. Guideline Steps

2.3.1. Protocol

A study protocol was developed to delineate the primary objectives, desired outcomes, proposed analytical methods, timelines, and deliverables for publication. The initial preparation and subsequent revisions of the protocol occurred over 6 months, whereas ethics and data sharing agreements took a further 6 months.

Due to the pilot nature of this study, the clinical networks involved were limited to the three listed above. An introductory letter was sent to respective Heads of Unit and following agreement individual data dictionaries were obtained. Features pertaining to diagnosis and treatment of MTC were identified and shortlisted as suitable for harmonization. From this, a reference template was devised and shared with the clinical networks. The reference template included 46 features and preferred formatting styles (see Terminology).

Non‐identifiable data were received from each clinical network via a secure file transfer process and assessed for duplicates, invalid values (characters, dates, other inconsistencies), and blank entries. Individual data items were manually graded based on similarity to the template (exact, close or low match) and the possibility of transformation without data loss. The harmonized dataset was developed using exact and closely matched data only.

2.3.2. Terminology

We refer to observations as the unit of analysis within a dataset (e.g. patient). Features are the individual measurable property being studied (e.g. age). Data points are the specific value for each feature (e.g. 47 years).

Data completeness was defined as the number of data points received compared with the number of data points expected.

Exact matches were identical in every detail: name of feature (e.g., basis of diagnosis) and format (e.g., 1, histology of primary tumor; 2, histology of metastases; 3, cytology; 99, unknown). Close matches were similar and could be transferred without data loss but were not identical. Low matches were highly variable and could not be transferred without considerable missing information.

2.3.3. Study Population

We included adult patients (> 18 years) with MTC as determined by histopathology, enrolled within one of the clinical networks between January 1, 2018 and December 31, 2021. There were no exclusions. As the data involved was secondary data, patient recruitment was not applicable. Informed consent for use of de‐identified data for research was obtained by the site‐specific clinical team at the time of initial enrollment and data entry. Ethical approval for this study was obtained by the Alfred Human Research Ethics Committee (706/22).

2.3.4. Outcome

We selected the presence of cervical lymph node metastases as the outcome of interest as an example to demonstrate the utility of harmonization, owing to its association with distant metastases and overall survival of patients with MTC. This was a binary outcome (yes/no) for all three networks.

2.3.5. Co‐Variates

We collapsed age into two categories (< 55 years and > 55 years) spanning the age range across networks (18–96 years). Age > 55 years is a high‐risk factor for recurrence for DTC [5, 11], although less convincingly for MTC. All networks used male and female categories for sex. Country of origin was either Australia or the United States. Several features had binary outcomes: vital status, hyperparathyroidism at presentation, diagnosis as an incidental finding, genetic testing, presence of palpable lymph nodes, abnormal lymph nodes on ultrasound, pre‐operative FDG‐PET, pre‐operative CT neck, fine needle aspiration, surgical treatment, MTC as the primary pathology, and confirmed lymph node metastases. The basis of diagnosis for MTC was determined either by cytology, or histology of the primary cancer, or histology of a metastasis. We collapsed neck examination into four relevant categories: normal, solitary nodule, multinodular goiter, and other. Thyroid cytology was consistent with the Bethesda classification (B1‐6), and remained relevant as a pre‐operative investigation before MTC was suspected or confirmed. We also collapsed size of primary tumor into two categories (< 20 mm and > 20 mm) since tumor size in MTC significantly affects prognosis, with larger tumors associated with a worse recurrence free survival [5, 8, 12]. The burden of lymph node disease was represented as lymph node ratio (LNR). This was defined as the number of lymph nodes involved with cancer compared with the number of lymph nodes excised. The LNR was collapsed into two groups (LNR< 0.4 and > 0.4), on the basis of its association with poor disease free survival [13, 14, 15].

2.3.6. Data Processing, Quality, and Analysis

We created a template to compare summary‐level results from each network. Feasibility of harmonization was defined as percent harmonizable (ratio of the sum of exact and closely matched data vs. all data received). Quality of harmonization was assessed by comparing descriptive statistics across levels of the harmonized variables, with consideration of differences between the networks. Continuous variables were summarized using the median and interquartile range, whereas mean and standard deviation, counts, and proportions were used for categorical variables.

A binary logistic regression analysis (backward stepwise likelihood ratio method) was performed using SPSS (Version 29) to determine risk factors for cervical lymph node metastases. The dependent variable was binary (present/absent) and was inclusive of all patients. Each of the clinical sites recorded cervical node metastases as a unique field, and did not derive these data secondarily from either N status or lymph node surgery. Co‐variates with limited data (e.g., < 5 observations) were consolidated to improve model stability and validity. Missing values were managed by assigning a temporary code (999) and specifying this value as missing in SPSS, or by simple imputation methods (replacing the missing value with the modal category (categorical variables) or median (numerical variables).

During initial modeling, SPSS automatically selected “Unknown” as the reference category for several multi‐level categorical variables because it appeared first alphabetically. This differed from the missing‐value handling described above and produced unstable comparisons due to the small number of observations coded as “Unknown.” To resolve this, all categorical variables were recoded so that a clinically meaningful and adequately populated category served as the reference level before rerunning the final regression model.

Model performance was evaluated using −2 Log Likelihood, Hosmer–Lemeshow goodness‐of‐fit, and Nagelkerke R². Final model coefficients were expressed as odds ratios (OR) with 95% confidence intervals (CI).

3. Results

3.1. Overview of Data

There were 114 patients over 4 years from 3 clinical networks and 17 sites. The completeness of data from the clinical networks ranged between 57.4% and 97.3%. One site shared partially cleansed data whereas the other two sites shared direct extracts from their databases. The percentage of features requested with an exact match ranged between 54% and 87% and approximately 13%–15% with a close match. Overall, 80.8% of data received were suitable for inclusion in the harmonized dataset. Table 1 shows a high level overview of the data.

TABLE 1.

High level overview of data.

	Network 1 (8 sites)	Network 2 (7 sites)	Network 3 (2 sites)	Harmonized dataset
Number of observations (O)	45	16	53	114
Number of features requested (F)	46	46	46	30
Number of data points expected (D = OxF)	2070	736	2438	5244
Number of features received (f)	35	32	46
Number of blank data points (b)	117	89	67
Number of data points received (d = D‐b)	1458	423	2371	4674
Completeness of data (C = d/D x100%)	70.4%	57.4%	97.3%	81.1%
Number of features with an exact match	28	25	40
% of available features	80%	78.1%	87.0%
% of requested features	60.9%	54.3%	87.0%
Number of features with a close match	7	6	6
% of available features	20%	18.8%	13.0%
% of requested features	15.2%	13.0%	13.0%
Number of features with a low match or unmatched	0	1	0
% of available features	0%	3.1%	0%
% of requested features	0%	2.2%	0%
Number of data points used for harmonization
Exact match	28 x 45 − 114 = 1146	25 x 16 − 89 = 311	40 x 53 − 67 = 2053	3510
Close match	7 x 45 − 3 = 312	6 x 16 − 0 = 96	6 x 53 − 0 = 318	726
Number of data points available in harmonized dataset	1458	407	2371	4236 (80.8%)

3.2. Harmonized Dataset

The harmonized dataset comprised 30 features which either had an exact or close match across all 17 sites. The mean age was 60.4 years and 54.4% were female. Just over half of the population were Australian patients (53.5%). 29.8% of patients had a normal thyroid on examination. 15.8% had palpable lymph nodes at presentation. Table 2 shows an overview of the harmonized data.

TABLE 2.

Overview of harmonized data.

	Registry 1 (n = 45)	Registry 2 (n = 16)	Registry 3 (n = 53)	Harmonized dataset (n = 114)
Age in years at diagnosis
Mean (SD)	63 (18.3)	60 (19.5)	58.7 (20.1)	60.4 (19.3)
Median (IQR)	65 (53.77)	57 (51.79)	60 (40.72)	63 (45.75)
Sex female (%)	20 (44.4)	10 (44.4)	32 (60.4)	62 (54.4)
Country (AUS)	45 (100)	16 (100)	0 (0)	61 (53.5)
Vital status alive	45 (100)	16 (100)	49 (92.5)	110 (96.5)
Basis of diagnosis
Histology of primary tumor	19 (42.2)	16 (100)	51 (96.2)	85 (74.6)
Histology of metastases	1 (2.2)	0 (0)	0 (0)	1 (0.9)
Cytology	25 (55.6)	0 (0)	0 (0)	26 (22.8)
Unknown	0 (0)	0 (0)	2 (3.8)	2 (1.8)
Incidental finding	7 (15.9)	6 (42.9)	24 (45.3)	37 (32.5)
Genetic testing	31 (70.1)	0 (0)	34 (64.2)	65 (57.0)
Hyperparathyroidism at first presentation	4 (9.3)	1 (6.3)	8 (15.1)	13 (11.4)
Neck examination
Normal	6 (14.3)	2 (15.4)	26 (49.1)	34 (29.8)
Single nodule	12 (28.6)	8 (61.5)	19 (35.8)	39 (34.2)
Multinodular goiter	19 (45.2)	3 (23.1)	3 (5.7)	25 (21.9)
Diffuse goiter	0 (0)	0 (0)	4 (7.5)	4 (3.5)
Other	2 (4.8)	0 (0)	0 (0)	2 (1.8)
Unknown	3 (6.7)	3 (1.9)	1 (1.9)	7 (6.1)
Palpable lymph nodes	8 (18.2)	4 (30.8)	6 (11.3)	18 (15.8)
Pre‐operative neck ultrasound	38 (84.4)	15 (93.8)	47 (88.7)	100 (87.7)
Abnormal lymph nodes on ultrasound	11 (29.7)	4 (25)	12 (25.5)	27 (23.7)
Pre‐operative FDG‐PET	12 (26.7)	1 (6.3)	3 (5.6)	16 (14.0)
Pre‐operative CT neck	17 (37.8)	7 (43.8)	17 (32.1)	41 (36.0)
Fine needle aspiration thyroid	36 (80)	13 (86.7)	31 (58.5)	80 (70.2)
Cytology of neck fine needle aspiration
Non‐diagnostic	1 (2.8)	1 (7.7)	2 (6.5)	4 (3.5)
Benign	2 (5.6)	0 (0)	1 (3.2)	3 (2.6)
Atypia of unknown significance	4 (11.1)	1 (7.7)	7 (22.6)	12 (10.5)
Follicular neoplasm	6 (16.7)	0 (0)	0 (0)	6 (5.2)
Suspicious for malignancy	1 (2.8)	1 (7.7)	5 (16.1)	7 (6.1)
Malignant	22 (61.1)	10 (77)	17 (54.8)	49 (43.0)
Site of first fine needle aspiration
Right lobe	15 (41.7)	7 (53.8)	15 (48.4)	37 (32.5)
Left lobe	20 (55.5)	6 (46.2)	10 (32.3)	36 (31.6)
Isthmus	0 (0)	0 (0)	0 (0)	0 (0)
Right lymph node	0 (0)	0 (0)	1 (3.2)	1 (0.8)
Left lymph node	0 (0)	0 (0)	4 (12.9)	4 (3.5)
Neck mass	1 (2.8)	0 (0)	0 (0)	1 (0.8)
Other	0 (0)	0 (0)	1 (3.2)	1 (0.8)
Surgery performed	45 (100)	16 (100)	53 (100)	114 (100)
Date of surgery
2018	11 (24.4)	1 (6.3)	13 (24.5)	25 (21.9)
2019	11 (24.4)	8 (50)	21 (39.6)	40 (35.1)
2020	11 (24.4)	2 (12.5)	9 (17.0)	22 (19.3)
2021	12 (26.7)	5 (31.3)	10 (17.0)	27 (23.7)
Extent of surgery
Hemi thyroid	10 (22.2)	2 (12.5)	4 (7.7)	16 (14.0)
Total thyroid	27 (60)	0 (0)	19 (36.5)	46 (40.3)
Total thyroid and central neck	0 (0)	6 (37.5)	20 (37.7)	26 (22.8)
Total thyroid, central and lateral neck	0 (0)	8 (50)	10 (19.2)	18 (15.8)
Central neck only	0 (0)	0 (0)	0 (0)	0 (0)
Lateral neck only	0 (0)	0 (0)	0 (0)	0 (0)
Other	8 (17.8)	0 (0)	0 (0)	8 (7.0)
Indication for surgery
Malignancy	28 (63.6)	14 (87.5)	27 (51.9)	69 (60.5)
Risk of malignancy	11 (25)	2 (12.5)	15 (28.8)	28 (24.6)
Compression	6 (13.3)	0 (0)	9 (16.9)	15 (13.2)
Hyperparathyroidism	0 (0)	0 (0)	2 (3.8)	2 (1.8)
Residual tumor after surgery
R0	30 (73.2)	13 (92.9)	45 (84.9)	88 (77.2)
R1	0 (0)	0 (0)	1 (1.9)	1 (0.8)
R2	4 (9.8)	0 (0)	5 (16.1)	9 (7.9)
RX	6 (14.6)	1 (7.1)	2 (6.5)	9 (7.9)
Other	1 (2.4)	0 (0)	0 (0)	1 (0.8)
Primary thyroid pathology MTC	45 (100)	16 (100)	51 (96.2)	112 (98.2)
Largest tumor deposit in mm
‐Mean (SD)	24 (19)	31 (14)	20.8 (20)	13 (20)
‐Median (IQR)	20 (11, 34)	31 (19.44)	16 (3.35)	6 (1.25)
Lymph node metastasis	28 (62.2)	10 (75)	20 (38.5)	58 (50.9)
Number of lymph nodes examined
Mean (SD)	18 (21)	18 (19)	13 (16)	16 (20)
Median (IQR)	9 (2, 23)	11 (6.29)	4 (1.13)	7 (1.23)
Number of positive lymph nodes
Mean (SD)	5 (9)	7 (9.5)	5.2 (8)	5 (11)
Median (IQR)	2 (0, 7.3)	3 (0.11.5)	0 (0.4)	1 (0.6)
T stage
T1a	9 (24.3)	1 (7.1)	25 (47.1)	34 (29.8)
T1b	9 (24.3)	3 (21.4)	7 (13.2)	19 (16.6)
T2	10 (27.0)	3 (21.4)	12 (22.6)	25 (21.9)
T3a	7 (18.9)	7 (50)	4 (7.5)	18 (15.8)
T3b	1 (2.7)	0 (0)	1 (1.9)	2 (1.8)
T4a	1 (2.7)	0 (0)	4 (7.5)	5 (4.4)
N stage
N0a	9 (20.9)	6 (40)	22 (41.5)	37 (32.5)
N0b	1 (2.3)	0 (0)	0 (0)	1 (0.8)
N1a	8 (18.6)	2 (13.3)	11 (20.8)	21 (18.4)
N1b	18 (41.9)	7 (46.7)	10 (18.9)	35 (30.7)
NX	2 (4.7)	0 (0)	10 (18.9)	12 (10.5)
M stage
M0	31 (77.5)	2 (12.5)	38 (71.7)	71 (62.3)
M1	4 (10)	0 (0)	4 (7.5)	8 (7.0)
MX	5 (12.5)	14 (87.5)	11 (20.8)	30 (26.3)

All patients had a pre‐operative ultrasound and 70.2% had a fine needle aspiration. The most common preoperative cytological diagnosis was malignancy/Bethesda 6 (43%). Almost one quarter (23.7%) had sonographically abnormal nodes.

The entire population was treated surgically, which occurred during the study period. Total thyroidectomy alone or in combination with a central neck dissection were the two most common procedures performed. 14% of patients were treated by hemithyroidectomy. The mean size of the largest tumor deposit was 13 mm. Of the patients who underwent a lymph node dissection (45.6%, 26/114 total thyroidectomy and central neck dissection; 18/114 total thyroidectomy, central neck and lateral neck dissection; and 8/114 other non‐compartmental lymph node dissection). The burden of lymph node disease, expressed as LNR (for all patients who underwent nodal resection), was 0.31 (31%). Overall, most patient's tumors were small and confined to the neck—the most common stage was T1a (28%), N0a (32.5%), M0 (62.3%).

Genetic testing was documented for 57% of the harmonized population. However, the presence of identifiable mutations was uncommon and if present was infrequently recorded. The co‐incidence of primary hyperparathyroidism was 11.4%.

3.3. Low Match/Unmatched

Five features could not be appropriately matched as the data were too heterogenous. This included year of death, results of genetic testing, history of recurrence, time to recurrence, and site of recurrence. Table 3 outlines the non‐harmonizable features.

TABLE 3.

Non‐harmonizable features.

Co‐variate
Year of death
Positive mutation identified
RET/PTC
MEN2
Familial
MEN2A
MEN2B

Recurrence
Years from first surgery to recurrence
Site of recurrence
Thyroid bed
Lateral neck
Other

Note: Re‐arranged during Transfection (RET), papillary thyroid cancer (PTC), Multiple Endocrine Neoplasia Type 2 (MEN2).

3.4. Regression Analysis

A binary logistic regression was performed to identify factors associated with the presence of cervical lymph node metastases. After removal of non‐significant predictors, three variables remained in the final model. Patients less than 55 years (OR = 3.68, 95% CI 1.21––11.24, p = 0.02) and patients with abnormal nodes on ultrasound (OR = 6.76, 95% CI 1.54–29.60, p = 0.01) had higher odds of cervical lymph node metastases. By contrast, there was an inverse relationship for patients with an incidental diagnosis of MTC (OR = 0.23, 95% CI 0.07–0.71, p = 0.01). No significant associations were observed for sex, size, basis of diagnosis, neck examination, cytology, indication for surgery, or individual registry. The final model demonstrated good calibration (Hosmer–Lemeshow p > 0.47) and explained more than half of the variance in status of cervical lymph node metastases (Nagelkerke R ² = 0.59). Table 4 shows the results of the final Logistic Regression Model for Palpable Lymph Nodes (Step 6a).

TABLE 4.

Final logistic regression model for palpable lymph nodes (step 6 ^a ).

Co‐variate	Category (reference)	Odds ratio	95% CI	p‐value
Age at diagnosis	< 55 years (ref = < 55 years)	3.68	1.21–11.24	0.02
Abnormal nodes on ultrasound	Yes (ref = no)	6.76	1.54–29.6	0.01
Incidental diagnosis	Yes (ref = no)	0.23	0.07–0.71	0.01

^a Variable(s) entered on step 1: Age at diagnosis, Sex, Basis of diagnosis, Incidental diagnosis, Neck exam, Abnormal Nodes on US, Cytology, Indication for Surgery, Size (mm), Registry.

4. Discussion

We have demonstrated feasibility of retrospective data mapping and harmonization using the Maelstrom framework. Overall, 80.8% of data received from three clinical networks were suitable for harmonization. Secondly, we have demonstrated clinical purpose of the harmonized dataset by interrogating predictors of clinically relevant endpoints, such as cervical lymph node metastases.

In the study of rare disease, such as MTC, research is often limited by low numbers which makes drawing meaningful conclusions difficult. The Maelstrom guidelines [9] were developed in 2017 by the Maelstrom Research team in Canada, with the aim of providing a systematic template as to how to combine retrospective data from multiple sources. They have previously been used to demonstrate harmonization of cohort studies and other clinical registries with success [16, 17, 18, 19]. The guidelines are strengthened by reproducibility, transparency, practicality, and support for large scale collaborations. By mandating detailed documentation of each step and comprehensive reporting of decision making, they ensure that harmonization processes are auditable. However, there are notable constraints, primarily stemming from the complexities of research involving real‐world data. The guidelines are not necessarily suitable for prospective collaboration and implementation can be resource intense. It is also noted that the guidelines have not been updated since inception and may be beginning to lag behind rapidly evolving tools in data science such as artificial intelligence.

In comparison with a systematic review and meta‐analysis which relies on aggregate data, researchers performing data mapping, and harmonization utilize raw data. Raw data are advantageous because coding errors and implausible values may be corrected and unpublished outcomes may be included, whereas with aggregate data authors are generally limited to what was included in the published manuscript. Furthermore, harmonization aligns all variables to the same common standard (e.g., age category–one study may define this as > 65 whereas another may define this as > 70), and the analysis is conducted with identical statistical methods [20, 21]. This reduces bias. Harmonization also reduces measurement heterogeneity. For example, different studies may measure the same concept in different ways. Although traditional meta‐analyses either note such variability as a limitation, or average it via a random‐effects model, studies that harmonize resolve heterogeneity by developing a common data model. In practice, systematic reviews and meta‐analyses are suitable when raw data are unavailable, for very large and well reported trials, and research questions focused on published literature (e.g. guideline development) [22]. However, increasingly major researchers and consortia now prioritize harmonization for pooled analyses.

Although not delineated as an individual step, the preparation and approval of network‐specific data sharing agreements (DSA) is also required and is distinct from local ethics applications. A DSA is a legally binding contract between two or more parties that outlines the terms for exchanging data. Typically, it defines how data can be used, should be protected, and must be shared (e.g., via a secure‐file transfer process). A robust DSA should include essential clauses which outline purpose, parties involved, lawful basis (legal grounds for sharing), data security, data subject rights, duration and termination, liability and indemnification, dispute resolution, and costs [23, 24, 25]. This is a complex and time‐consuming step, particularly for international sites, which have different legal requirements.

In our harmonized cohort, the most common stage of MTC was T1aN0. It would seem plausible that this is secondary to a high rate of pre‐symptomatic screening. Unfortunately, this is not possible to confirm as the completeness of genetic testing and documentation of results was low. We suspect this is related to the timing of data entry but also specific restrictions within the DSA. Typically, the most common stage of MTC at the time of diagnosis and treatment is stage IV [26]. Early detection is uncommon unless there is a known familial syndrome, and the biology of disease is characterized by early metastases to regional lymph nodes. Our results are unexpected and warrant further investigation.

We performed a regression analysis using the harmonized dataset to determine risk factors for cervical lymph node metastases. Cervical lymph node metastases are a poor prognostic factor in MTC, and are directly associated with the presence of distant metastases and inferior cancer‐specific survival [26]. In our analysis, three variables emerged as contributors in the final model: age < 55 years, sonographically abnormal lymph nodes, and incidental presentation; however, only age was clinically meaningful. These findings have previously been suggested as risk factors for lymph node metastases in MTC. Zhu et al. reported age < 52 years and sonographic irregularity as predictors of cervical lymph node metastases [27], whereas Wang et al. demonstrated a positive association with symptomatic presentations (inverse association with incidental presentation) [28]. Of note, a recent systematic review and meta‐analysis focused on predominantly histological features (e.g., tumor size > 1 cm, multifocality, bilaterality, capsular invasion, and extrathyroidal extension) [29]. However, since the lymph node dissection is often performed concurrently with resection of the primary tumor [30], the utility of histological red flags is not as logical to the authors. Other reported risk factors for lymph node metastases in MTC include high pre‐operative calcitonin and CEA, central lymph node metastases, and male sex [31, 32]. Ideally, future studies would focus on factors which are identifiable pre‐operatively and can therefore be used to counsel the patient before treatment commences.

The harmonization process was critical to the success of this analysis. Harmonization allowed data to be pooled, reducing fragmentation and data loss. It also reduced missingness by aligning fields across networks so that variables present in one dataset may inform equivalent fields in another. Despite the modest sample size, our final model achieved good calibration and explained more than half of the variability in status of cervical lymph node metastases. Wide confidence intervals reflect limited subgroup counts rather than absence of association.

Points of difference with this study are the use of the Maelstrom guidelines for endocrine surgical pathology, international collaboration, and the opportunity for future projects on a larger scale. Compared with medicine, surgical data are commonly retrospective. In the most recent ATA guidelines for thyroid cancer, only 52 out of 934 references refer to prospective cohorts [10]. This manuscript demonstrates how the Maelstrom guidelines can be used to harmonize surgical data which are frequently retrospective. For clinicians investigating rare disease this is particularly valid, whereby the primary aim is to generate meaningful results. To improve understanding of disease and patient outcome. Data mapping and harmonization (enabling individual participant data meta‐analysis) make the shift from summarizing published reports to analyzing what occurred at an individual level [20].

There are a few limitations of this study. First, not all data are harmonizable due to heterogeneity in collection methods and ethical constraints. In our study, we could not harmonize results from genetic testing which was particularly unfortunate for a population of patients with MTC. One site was legally prevented from sharing this information, if the specific mutation was present in less than three other patients, in order to maintain absolute anonymity. For other sites, we suspect genetic testing was performed, but the results were not recorded in the databases (as they were not available at the time of data entry). Certainly, in Australia, there is considerable delay between test requested, performed, and results being available. Unfortunately, we are unable to verify this due to the study design and this is an ongoing limitation of research using secondary sources.

We were also unable to harmonize certain clinically relevant outcomes such as death or recurrence due to large scale missingness. In most circumstances, blank entries probably meant a clinical event had not occurred but this could not be confirmed. Of note, we observed that the network with the highest percentage of data completeness uses the electronic medical record as a surrogate registry and is not reliant on manual, often duplicate, data entry into a standalone database. This is a consideration for future projects. Ultimately, the quality of the harmonized dataset is proportional to the quality of the individual datasets, but is superior to any of the datasets in isolation.

Secondly, our study period is relatively short and as such absolute patient numbers are low. This limits the significance of our results. We concede this was intentional, allowing us to focus on the primary aim, to demonstrate the mechanics of data mapping and harmonization, and complete the project in a reasonable timeframe. We encourage future projects, involving additional sites, more data fields, and longer periods of investigation, to investigate clinically relevant outcome measures using the same methodology as we have outlined.

Thirdly, harmonization is time consuming, especially the index interaction with a clinical network. In our experience, the DSA was the most laborious component. However, once DSA is established, future collaborations should be easier. We learned that it is helpful to have a co‐ordinating clinician at each site and to communicate with the team regularly, especially when there are lengthy delays. We also observed that for sites with aligning data dictionaries mapping is intuitively more streamlined. This is a consideration in the design of new data repositories or in choosing with whom to collaborate. As the scope and integrity of artificial intelligence continues to evolve, efficiency of data mapping and harmonization should significantly improve.

The aim of this paper was to showcase the mechanics of data mapping and harmonization among clinical sites investigating MTC internationally. The purpose of the logistic regression analysis was to give clinical relevance to the harmonized dataset. This paper demonstrates how to handle small volume, heterogenous, raw data. This is particularly relevant to researchers of rare disease or obscure endpoints. We hope that future studies can build on our methodology. Inclusion of recurrence or survival data and additional sites would add value.

5. Conclusion

Data mapping and harmonization among clinical networks is feasible. The Maelstrom guidelines provide a methodological template to achieve this efficiently. In this harmonized multi‐registry MTC dataset, the odds of cervical metastases were independently associated with younger patient age and sonographically abnormal nodes. There was an inverse association with an incidental presentation. Data harmonization enabled pooling of raw data from multiple sources and meaningful modeling that would not have been possible for any dataset in isolation. Data mapping and harmonization are useful tools for researchers investigating rare diseases.

Author Contributions

Edwina C. Moore: conceptualization, writing – original draft, writing – review and editing, data curation, methodology, formal analysis, project administration. Jonathan Serpell: validation, writing – review and editing, supervision. Rasa Ruseckaite: supervision, writing – review and editing. Liane Ioannou: supervision, writing – review and editing. Justin Bauzon: data curation, writing – review and editing. Gustavo Romero‐Velez: data curation, writing – review and editing. Joyce Shin: data curation. Allan Siperstein: data curation. Alex Papachristos: data curation. Stan Sidhu: data curation. Mark Sywak: data curation. Susannah Ahern: methodology, resources, supervision, validation, writing – review and editing. Ahmad Pourghaderi: methodology, formal analysis.

Funding

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Moore E. C., Serpell J., Ruseckaite R., et al., “Data Harmonization for Collaborative Research Among Australian and US Registries: A Case Study in Medullary Thyroid Cancer (MTC),” World Journal of Surgery (2026): 50. no. 7), 1940–1949, 10.1002/wjs.70436.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.