The first task of the file processing was to find and remove multiple copies of the same file; such duplicates can happen, as the same file can be harvested from multiple users. For each pair of files, we compared their sizes (in kilobytes), titles, and first and last sentences. If all the parameters were the same, we tagged two files as duplicates and kept only one for further processing.

We assumed that any published text was not leaking PHI—for example, writings describing fictional characters, and magazine and newspaper articles that contain information that is already public. We used Amazon Web Services as a source database of publication titles [25]. Although the number of titles fluctuates almost daily, the database has 400,000 to 500,000 titles for books; recording companies such as Sony, EMI, and Universal have more than 250,000 music titles in the database. Files with exact matching titles were discarded. Exceptions were made for files with titles that included such words as notification, affidavit, justice, discharge, and lab. Our system did not discard these files and retained them for further processing. If there was no exact title matching, the file was passed on for further processing.

The text extractor converted files into raw text by removing all formatting meta-information. It also discarded nontext files (eg, images and music), corrupted files, and viruses. A wide range of input file formats suggested the use of several tools, with each tool extracting text from specific formats. We used the open source program Antiword to extract text from Microsoft Word documents [26]. To extract text from PDF, RTF, HTML, or XML files, we used the open source programs MineText [27] and GetText [28].

We differentiated between English-language texts and texts written in other languages. We applied a publicly available language identifier, TextCat Language Guesser, which can identify 69 languages [29]. For text, the tool outputs several possible languages. If English was the most likely language of the text, then it appeared at the beginning of the output. Our manual examination had shown that, in our sample data, the first English tag always correctly marked texts written in English. We discarded a file if English was not the most likely language of the text.

P2P Watch looked for files with nonpersonal content. It filtered out published and educational materials (eg, assignments and theses) and nonpersonal texts (eg, manuals and technical reports) that were not found by the title lookup. We also hypothesized that music lyrics, discussion of popular fictional characters or current political events, and advertisement would be unlikely candidates for leaking explicit, detailed PHI. We built a list of fictional characters (eg, Bart Simpson), celebrities (eg, Paris Hilton), and public figures (eg, George Bush). We considered that, by the nature of their occupations, celebrities and public figures would have a lot of information about them publicly known and therefore any PHI pertaining to these individuals would not be considered a breach. To perform this task, we built a list of terms that appeared in the preface of publishable and educational texts; the terms are listed in Multimedia Appendix 1. We used regular expressions to locate those terms and their variations in the first 200 words of the file. Table 1 lists word categories and examples.

We considered that a person can be identified from first and last names, addresses, dates (which can be linked to identifiable events such as birth, death, and marriage), and organization names (eg, school, church, and professional association). We divided PII into three categories: person information (eg, names, family relations, and age-defining events), structured information (eg, dates, telephone numbers, and email), and geographic location (eg, street address and organization). For instance, telephone numbers are both geographic identifiers and numeric identifiers. Figure 2 illustrates the category relations.

For the personal name lookup, we acquired a female first name list, a male first name list, and a last name list. The lists contained both formal and informal name forms (eg, William, Bill, and Billy) and non-Anglo-Saxon names (eg, Meehai and Leila), as well as common misspellings (eg, Bll). The lists came with a commercial database marketing tool that performed best in an independent evaluation [30].

To reduce computationally expensive person name lookups, we first searched for patterns of family relations in the text (eg, “my daughter” and “an uncle of”), self-identification (eg, “my name” and “sincerely”), or life event (eg, “was born” and “died in”). Depending on the patterns, either the preceding or the following capitalized words were stored in a file’s name list. Having a file’s name list considerably accelerated the file processing. Further, when P2P Watch checked for a person name, it first checked with the file’s name list.

Additionally, PII included standardized information as follows: (1) telephone numbers: we looked for complete and incomplete formats, used in North America, (2) health insurance numbers: we looked for health insurance numbers assigned by each province (Canada), (3) dates: we restricted the dates to the 20th and 21st centuries, as earlier dates are unlikely to be related to health information of living human beings; also, dates had to be specific: “March 9th, 1999” indicated a specific date, whereas “March was chilly” did not, (4) email address: for example, john@canada.ca, john AT Canada DOT ca, (5) postal codes in Canada and zip code in the United States. These five categories were retrieved by manually built soft regular expressions.

Apart from geographic locations in Canada and the United States, we used some international information and introduced different granularity for different geographic categories. First was country: all the UN-recognized countries and their capitals (eg, France and Paris; Liberia and Monrovia) and self-proclaimed entities (eg, Eritrea and Abkhazia). Second was place. In the United States, we used state name, state capital, and—to cover the biggest single population unit—the largest city in the state; for example, we had Illinois, Springfield, and Chicago. In Canada, we used province, provincial capital, largest cities, and tourist attractions (eg, Alberta, Edmonton, Calgary, Banff), and the same for territories. In Europe, Latin America, Asia, Africa, and Australia, we used major cities that are not national capitals. The list is given in Multimedia Appendix 2.

The Canadian Judicial Council further considers certain organizations to be part of PII [31], such as public institutions (eg, schools and churches), care providers (eg, aid societies and foster homes), the names of support organizations (eg, women’s and senior’s support centers), and other location identifiers (eg, educational institutions and military bases). The organization names were expressed in many language forms. We modeled language patterns (eg, “lived in” and “come from”), organization types (eg, schools, military units, and churches), and the target population (eg, youth, women, and seniors).

To be marked as a text with PII, the file had to contain a geographic identifier and two other personal identifiers, such as a person’s first name and last name, or a person’s first name and date of birth. All the files marked as PII were passed on to the last component.

Disease names (eg, arthritis and mumps) and symptoms (eg, chest pain and headache), and procedures (eg, heart surgery and x-rays) most directly convey health information that is usually discussed in a clinical setting. To build a list of corresponding terms, we used the International Classification of Diseases [32] and the Medical Dictionary for Regulatory Activities (MedDRA) [33]. Drug names, too, may allow one to infer a specific medical, behavioral, or psychological condition or ailment of another individual. We used the Canadian Drug Product Database (Active and Inactive), which contains the names of drugs approved for use in Canada and previously available drugs [34]. To accommodate extraction of various drug names, we obtained a list of generic drug names and the trade names associated with them [35]. However, the resources listed above leave some gaps in the detection of health information. The most noticeable are acronyms (eg, ICU), specialties of health care providers (therapist, surgeon), and some condition names (blood pressure, tube fed). To fill the gaps, we manually searched Webster’s New World Medical Dictionary [36]. We minimized the above-mentioned resources by removing unrelated categories (animal diseases, animal drugs). Then the remaining texts were converted to lowercase, punctuation marks and numbers were removed, and stop words (eg, of and when) were eliminated. The list of the resulting keywords is given in Multimedia Appendix 3.

More details on the method can be found in Sokolova et al [19].

The files were gathered from April 2008 to June 2009 from three networks. We selected FastTrack, Gnutella, and eD2K networks due to their global popularity and high share of users [2]. To automatically search for and download P2P files, we modified the publicly available Shareaza P2P client [37], which is a software package allowing one to connect to multiple P2P networks simultaneously. In-house modifications to Shareaza included changes to the search function and an increase of logging capabilities.

We focused on capturing the most popular document formats: Microsoft Word (.doc), raw text (.txt), Rich Text Format (.rtf), Excel (.xls), PowerPoint (.ppt), Portable Document Format (.pdf), Extensible Markup Language (.xml), and HyperText Markup Language (.html). The search function was modified to automatically search for those formats and automatically retrieve the files. Automatic searches were conducted by the code at 15 minute intervals.

In total, we have gathered 3924 files. The data were sent for processing as is, without preliminary normalization: we preserved all the initial spelling, capitalization, grammar, and so on.

We considerably reduced processing time by first testing whether a file was a duplicate of an already gathered file. Removal of duplicates does not affect the quality of PHI detection, as P2P Watch should process the original file. We discarded 514 files (eg, Quick Recipe And Meal Ideas, 36 Christmas Carols & Songs), mostly books and technical manuals; the size of discarded files was 786 MB.

We discarded 286 files, which contained content that was found in the Amazon.com database (eg, Iron Crypt of the Heretics; The Haunted Lighthouse; New York, New York; ABBA’s The Winner Takes It All; and The Abominable Snowman); the size of discarded files was 1234 MB. We requested the exact match of the file title with the database entry; hence, some files with published content were passed through this filter.

We worked with different file formats to extract text and discard images, virus files, and corrupted files. At this stage, 673 files were discarded, mostly mp3 and rm files with music and video contents; the size of the discarded files was 3077 MB.

We identified English-language texts among the remaining files; 882 non-English texts were discarded, with a total size of 10 MB. To avoid misrepresentation of North America as a polyglot continent, we checked the discarded files and found that many of them had explicitly erotic content.

Based on processing of the first 200 words, 880 files were discarded, with a total size of 171 MB. Books composed the majority of the discarded files, with the remaining part comprising cover letters, resumes, homework, and forms.

Based on the search for PII within a complete file, 550 files were discarded. Most of these texts were in a small publishable form such as articles, opinion pieces, local community letters, forms, and job application packages. The size of discarded files was 3.9 MB.

The last component identified 57 files as potential PHI and discarded 240 files. Among the discarded files, most promoted different types of consumer goods and services (credit services, fitness and skin care, gadgets, etc); some tax and insurance forms were discarded as well.