Ethical approval was obtained by the Area Vasta Emilia Centro Ethical Committee on December 12, 2019 (protocol number: 197/2018/Oss/AUSLF). This study conforms to the principles expressed in the Declaration of Helsinki.

In Italy, mental health care is provided by departments of mental health [16-19]. The levels of care within each department of mental health include community-based mental health centers, hospital psychiatric inpatient units, and rehabilitation or residential facilities. Each community-based mental health center serves as a hub of psychiatric care for geographically defined catchment areas with 50,000 to 150,000 inhabitants [20] In Ferrara, Northern Italy, the Integrated Department of Mental Health and Pathological Addiction covers an area of 2630 km^², with a catchment of 342,061 inhabitants as of 2020 [21].

Data were collected in 2 periods that were distinct in terms of the methodology used, the psychiatric services delivered, and the level of digitalization. Data related to the first period, which began in 1925 and ended in 1990, were gathered mostly in a psychiatric asylum, during a time when digital health was not fully developed or adopted.

In 1991, the first structured EHR (ie, SIPER [Sistema Informativo Psichiatrico dell'Emilia-Romagna]) was introduced and implemented locally by the Local Health Trust of Ferrara for Mental Health in Adults. Different software programs were adopted during the years following the implementation of SIPER, and each new software program replaced the previous one by importing already existing data and adding new features (and thus information), as detailed in Textbox 1.

The first goal was the creation of a new, fully deidentified database with data available in EFESO (Newteam SRL)—a necessary step for complying with privacy constraints.

In order to remove all protected health information (PHI), source data needed to be modified. This could not be done by directly editing data in EFESO, since the source could not be altered directly. Thus, the new research database—the Ferrara-Psychiatry (FEPSY) database—was built via the Extract, Transform, Load process, which is a 3-phase process [22] in which data are first extracted from 1 source or multiple and possibly different sources (eg, databases, flat or formatted files, and web pages). Afterward, the extracted data are stored in a staging area, where they undergo transformation, such as filtering, cleaning, summarization, and normalization. Finally, the transformed data are loaded into the destination storage. For example, one type of transformation was record exclusion. We excluded records of patients that could not be unequivocally identified by the tax code—a unique 16-digit alphanumeric code that identifies a person in Italian public administration forms.

While assessing the suitability of the FEPSY database for research purposes, we noted that historical information dating back up to 1925 had also been maintained in EFESO. We understood that such data were manually imported into the electronic databases that preceded SIPER (year 1991); however, because we could not confirm the procedures, scope, and quality of this historical data import, we decided to document the existence of these data but exclude them from analysis.

More details about the FEPSY database can be found in Multimedia Appendix 1. In Table S1 in Multimedia Appendix 1, for each table in the FEPSY database, the total number of rows (number and percentage of records retained in the FEPSY database) is reported, alongside the number of records in the corresponding original EFESO table from which data were extracted (number of records in the corresponding EFESO table). As detailed in Table S1 in Multimedia Appendix 1, of the 4,264,954 records, 3,861,432 (90.54%) were kept. These records included detailed information about the patient, their illness, and the treatments provided.

In Table S2 in Multimedia Appendix 1, 2 types of anomalies for each table in the FEPSY database are described; one is date inconsistency (eg, when the closing date precedes the opening date of the medical chart), and the other is a date anomaly that was generated by the automatized mechanism that was introduced by EFESO when migrating data from IPPOCRATE (GPI SRL; August 26, 2008).

Once the anonymized database was built, a clustering analysis was performed to investigate the data set quality. A clustering algorithm is an unsupervised machine learning technique that is used to group objects, so that objects of the same group (or cluster) are very similar to one another and objects of different groups are very dissimilar. To decide the degree of similarity (or dissimilarity) between 2 objects, various distance measures can be used, such as the Euclidean distance between (normalized) numerical representations of the objects. We tested the hypothesis that the patients modeled in the data set could be divided into homogeneous clusters. The k-means algorithm computes a numerical distance between objects to determine to which cluster they belong. However, in our case, data were categorical. In 1995, Ralambondrainy [23] introduced an approach that enables the use of the k-means algorithm with categorical data. In this approach, nominal attributes are converted into binary attributes—one for each value that the attribute can take—so that they can be considered as numerical attributes by the algorithm.

We performed the clustering analysis with the WEKA (Waikato Environment for Knowledge Analysis; University of Waikato) data mining tool, which provides an implementation of the k-means algorithm (ie, SimpleKMeans) and can handle categorical data [24]. SimpleKMeans can also handle missing values by replacing them with the mean or mode.

For this preliminary clustering analysis, we included only the patients who had at least one recorded diagnosis of a mental disorder (ie, International Classification of Diseases, Ninth Revision [ICD-9] codes 290-319) [25].

Patients were excluded if they had nonpsychiatric diagnoses (ICD-9 codes V01-V91; 2707/46,222, 5.86%) or had never received an ICD-9 diagnosis (5493/46,222, 11.88%). The resulting subset for the clustering analysis included 38,022 patients.

We considered sociodemographic variables, such as biological sex, age at first visit, nationality, marital status, living situation, education, occupational status, birthplace (district), and the catchment area (district) providing care (determined by domicile postal code or by residence postal code when the domicile was missing).

The sample included 46,222 individuals, whose sociodemographic characteristics are detailed in Table 1.

Data extracted from EFESO were loaded in the FEPSY database—the newly created MySQL (Oracle Corporation) relational database—by using the same automated procedure that was used to extract them. For each built-in table, an insert query, which took the values from the same DataFrame of the select query, was executed.

The data included in the final composite FEPSY data set were those collected from 1991 to February 2021. Since 1991, each year, a median of 1404 (IQR 1117.5-1757.7) individuals had newly accessed care, and a median of 7300 (IQR 6109.5-9397.5) individuals were actively receiving care, as represented in Figure 1. Figure 2 shows the number of patients treated per year in total and by sex. The sudden decrease observed in 2009 was due to an automated closing procedure that was introduced in 2008 and retained from then on. When migrating from IPPOCRATE to EFESO, all medical records, products, and diagnoses that had not been updated in 365 days were assumed to be closed or terminated, and the missing closing date was replaced with the date of the migration to EFESO, that is, August 26, 2008.

As described in Table 2, the most frequent diagnoses at first admission were depression and anxiety disorder. During the 30-year time span, more than half (32,230/46,222, 69.73%) of the patients had only 1 chart open, and only 5184 patients had at least one psychiatric hospitalization.

This analysis, which was carried out on the subset of 38,022 individuals who had at least one mental disorder diagnosis, identified 2 distinct clusters (Figure 3). One is represented by single male patients who were born in Ferrara, those who were living with parents, and those whose age at first visit was between 25 and 34 years; the other is represented by married female patients who were living with their own acquired families, those who were born outside the province of Ferrara, and those whose age at first visit was 35 to 44 years. The following sociodemographic features were similar in the two clusters: Italian nationality, individuals with a high school degree, employed individuals, and individuals who were receiving treatment in the Ferrara catchment area.

This study describes the process of adapting one of the longest running EHRs of public mental health care for research purposes. The FEPSY data set covers a catchment area with 342,061 inhabitants (as of 2020) and includes a total of 46,222 unique individuals who had access to mental health services over a span of 30 years (1991-2021). The FEPSY database is suitable for descriptive, predictive, and inferential analyses via conventional analysis and AI techniques, as demonstrated by the preliminary findings of the clustering analysis. To our knowledge, our database is the first of its kind in Italy. In Europe, longitudinal and prospective registries have long been in use. For example, large data sets were extracted from the Danish National Patient Registry [30] and the Danish Psychiatric Central Research Register [31]. The first data set contained data on 8,085,603 patients, which were collected from 1977 to 2012. The second data set included data on a total of 747,176 patients, which were collected from 1970 to 2010. In both cases, the register contained dates of the onset and end of any treatment, diagnoses, types of referrals, and places of treatment, thereby allowing for the possibility to perform health registry–based research [32], considering a total population of approximately 27 million.

The main finding of this study is that data that were not originally conceived for research were successfully extracted from EHR software and loaded into a new anonymized database. This step is of foremost importance, as the data originated from an information system that was changed and updated multiple times and was not designed to allow for exploratory investigations in a structured manner. Thus, this new data set may represent the ideal setting to build, test, and refine an analytical methodology for extracting data and preparing these data for research purposes. This methodology could also be applied to other clinical data sets, such as data sets from other medical disciplines (eg, oncology), with characteristics that are similar to those of the FEPSY data set [33]. Additionally, it will be of foremost importance to validate the methodological approach and findings of upcoming research originating from the FEPSY data set by proactively seeking collaboration with other research groups, in order to enable the replication of findings from the territory of Ferrara and the use of the FEPSY data set to replicate findings from research involving other registries.

This work also allows for collaborations in terms of learning health networks, which use comparable data that originate from EHRs to support clinical decisions, improve the delivery of efficient and effective medical care, and help with the integration of research in health care [34-36].

Our results will also pave the way for an in-depth study on the use of health care resources; the results will be used to develop a system that is capable of planning the use of such resources. Such a system would optimize the use of health care resources while maintaining or possibly improving the quality of treatment. For example, in Italy, Donisi et al [37] predicted the cost of community mental health care by using clinical and sociodemographic information originating from the Psychiatric Case Register in the Verona Health District. This allowed for the linking of social deprivation to psychiatric service utilization [38] and shed a light on possible contributors to social isolation in an already vulnerable population. Our clustering analysis, which was conducted on the FEPSY data set to test its feasibility and robustness, identified 2 clusters; women appeared to access mental health services later in life and were typically married, in comparison to men. These findings were consistent with the literature [39-41] and supported the ability of the extracted data to detect known patterns, even though the results should be interpreted with caution, given the large amount of missing sociodemographic data. Clustering analyses can be useful for building prediction models and planning a department's resource allocation, as they provide relevant information on patients at presentation and on illness trajectory [42].

The process described in this paper faced 5 major challenges that we mitigated, as follows. First, the source data included several built-in structural informatic elements (so-called tables) that had to be screened and deleted in order to get to the core data. Second, the anonymization step was of absolute importance, and in order to both comply with privacy constraints and be able to preserve the integrity of the data, the study team decided to keep only subelements of certain data items (eg, for the birth date, only the year was kept, and for residence, only the postal code was kept). Third, in order to establish which records were correct, an iterative comparison of the FEPSY database and the local and regional database was performed by a third party who had access to PHI. Fourth, records that were deemed unreliable were excluded (eg, clinical procedures referring to nonexistent medical records). Furthermore, the data extracted from EFESO originated from different software and thus possibly generated some errors. In the FEPSY data set, these errors seem to be limited to records, and the proportion of records with errors was very low (49,854/3,861,432, 1.29%). In the end, we decided to exclude data that were collected before 1991 and duplicate patient records. Fifth, missing data challenges were also addressed, especially in the clustering process. For this purpose, the WEKA data mining tool was used.

The quality and completeness of the collected and cleaned data, as well as the large number of records stored in the FEPSY database, resulted in the definition of a data set that is particularly suitable for automatic analysis and has appealing characteristics for research, such as a long period of data availability, great diversity in the sociodemographic factors of the patients represented, and a history of treatments and drugs administered. This could possibly represent a strong foundation for many different studies of mental illness and resource use, favoring comparisons between Italy and other countries regarding the delivery and quality of community and hospital psychiatric care [43-45]. Furthermore, the newly created database does not include sensitive information, even though this information can be retrieved by using an external supporting table created ad hoc, which could link the FEPSY data set with other data sets (eg, hospital data and tumor registries) for future research.

The novelty of this project is represented by its interdisciplinary nature (psychiatry, public health, epidemiology, sociology, mathematics, computer science, and AI), the potential versatility of the methods that can be used with the FEPSY database, and the versatility of the systems that could be created via analyses involving the FEPSY database. To our knowledge, this study is the first attempt to retrospectively build a single data set that includes more than 30 years’ worth of data on mental health services in a specific area.

Such a data set would also allow for longitudinal analyses, such as those that have already been performed with the Nordic registry (a prospective registry) and, more recently, the South London and Maudsley National Health Service [46] and the Camden & Islington Research Database [47].

We believe that historical data can add value to subsequent analyses, because they allow researchers to understand how mental health services have evolved over the past decades and the extent to which phenotypical presentations of different diseases have changed over time. In light of these considerations, factors that should be taken into account are (1) potential cohort and time effects, such as historical events (eg, the Great Recession in 2008); (2) changes in legal and medical approaches to mental health; and (3) changes in the classification of mental disorders [48].

Our findings must be interpreted in the light of some limitations. First, the sample size is limited by the geographical catchment. A larger catchment or a more densely populated region would probably have a larger volume of treated individuals and thus have more data, which would facilitate machine learning analyses. However, we believe that even if the sample is limited by the geographical catchment, the diverse socioeconomic distribution in Ferrara is a strength that mitigates this limitation, providing insight into the possible moderator or mediator roles of socioeconomic variables that are considered social determinants of mental health. Second, another potential limitation is the missing data for some sociodemographic attributes, which may reduce the statistical power of a study [49] or affect the accuracy of machine learning algorithms [50]. In order to overcome this issue, missing values can be handled with multiple imputation methods or replaced with the mean or the mode (ie, for quantitative or qualitative data, respectively). Moreover, sociodemographic information can be drawn from external and publicly available sources, such as the Italian National Institute of Statistics [51], which includes the census of the population as well as social, economic, and environmental surveys and analyses. Lastly, there is the risk of introducing bias while building prediction models, especially when using supervised machine learning techniques, due to small sample sizes and the poor handling of missing data and overfitting [52]. With regard to the sample size, our sample appears to be sufficiently large for risk prediction analyses. Overfitting can be addressed with penalized models [53].

This work sets a starting point for future investigations, which can be described as follows: (1) identifying patients who have a higher severity index or chronicity level and those who require a greater use of health resources; (2) identifying and validating, by means of machine learning models, demographic, clinical, and social predictors of clinically relevant outcomes that are useful for an ad hoc programming of resources (eg, sex, gender, or social deprivation [38,39,54]); (3) further optimizing and tailoring the analysis methods, so that they can also be applied to other data sets (eg, the local mental health registries for child and adolescent neuropsychiatry and for drug addiction services); (4) interacting with international learning health networks [55-57]; and (5) linking the FEPSY data set with external data sources, such as census data, tumor registries [58], death registries, and criminal justice data [59]. As a result of the increasing digitalization of medical records, it was possible to gather years of mental health history for every patient. This will enable for the conduct of symbolic and subsymbolic analyses on time series via automatic methodologies. Classical supervised and unsupervised machine learning and deep learning techniques will be evaluated. In order to explore the relationship between sociodemographic characteristics and specific diagnostic questions (eg, the incidence and prevalence of psychosis), a supervised framework will be deployed, in which binary labels (eg, “psychosis” or “no-psychosis”) or multiple classification labels (eg, “ICD-9 diagnosis”) will be associated with the patient. The main problem to overcome will be the imbalance of the data set, that is, when there is an unequal distribution of classes in the data set. In such instances, a standard machine learning technique, such as a support vector machine or random forest [60,61], will be applied. Moreover, each patient could potentially be considered as a distinct time series by including the temporal dimension of the treatment and by applying recurrent neural networks [62,63]. By doing so, the prediction of the new onset of a disease and the subsequent use of health resources will be the focus, in order to plan and optimize health care resources.

The process described in this study resulted in the building of a data set that included the information of 46,222 individuals who had access to psychiatric services in the Ferrara province over the course of almost 30 years. The preliminary findings from the clustering analysis confirmed the quality of the newly established database. The process we implemented proved to be a solid method that can be replicated with similar data sets, even if they were not originally compiled for research purposes.