Mir Siadaty on Biomedical Information Retrieval

The central feature of BioMedLib is that the user finds “better answers”, and finds them “faster”, as compared to other search engines.

How does BioMedLib find better answers than other engines?

1. BioMedLib performs a comprehensive analysis of your query. It is like mapping the query to MeSH terms or other controlled vocabularies used by search engines like Ovid/Embase/Pubmed, but 100 times more comprehensive. For example, MeSH has 25 thousand terms (concepts), while BioMedLib’s knowledgebase contains over 2 million biomedical concepts, expressed by 20 million unique terms (synonyms). There are many advantages to a better more comprehensive query mapping and expansion; like the user will miss less results due to user’s choice of words.

2. BioMedLib tags every instance of 20 million terms (synonyms) that express 2 million biomedical concepts, in each of the 20 million articles of Medline database. This is similar to MeSH terms added to each article by the NLM indexers, but much more comprehensive. BioMedLib uses optimized and automated algorithms that assign the tags without need for human intervention. This automation makes it possbile to retag the whole corpus of the 20 million articles about twice per year, when major vocabularies used in BioMedLib’s knowledgebase are updated.
Given MeSH, Mtree and other vocabularies (used by other search engines) improve with each annual edition, when was the last time Ovid/Embase/Pubmed retagged their whole corpus of articles?

3. BioMedLib computes an advanced hybrid relevance score, proprietary to BML, for each article of each query that you submit. By default BioMedLib shows the most relevant articles first, where they have the highest chance to be the “best answers” to your query.
Moreover, you can specify a range of publication dates, within which the articles are then sorted by relevance, thus providing you with both timeliness and relevance.

The factors that are used in computing BioMedLib’s relevance score include:

3.1 Semantics: presence of relation in addition to the presence of query words. BioMedLib looks for triplets {Concept1, Relation, Concept2} in each sentence of each article, where Concept1 and Concept2 are from your query. BioMedLib assigns higher relevance score to an article that contains such relation triplets.

3.2 Meaning-based (concept-based) search, besides text-word searching. When you submit a query like ‘heart attack’, BioMedLib not only searches for the words ‘heart’ and ‘attack’, but it also searches for the biomedical concept with ID = C0027051, which is a “Disease or Syndrome”, with definition of “gross necrosis of the myocardium, as a result of interruption of the blood supply to the area”, which has about 100 synonyms taken from about 150 different vocabularies. Other engines may do similar operation, but they understand only a small fraction of the 2 million concepts that BioMedLib understands. BioMedLib assigns higher relevance score to an article that contains the concept (than an article that contains the query words but in separate and unrelated places).

3.3 BioMedLib’s relevance score includes all the factors used in a typical TF-IDF (term frequency – inverse document frequency) operation. For example, field-weighting is one of such features, like when “major” MeSH terms of an article matches your query, versus the minor MeSHes, and that article is given a higher relevance score.

Therefore BioMedLib utilizes an optimized hybrid of multiple scoring systems, some unique to BML, to deliver its superior relevance score.

How do users find the best answers faster with BioMedLib?

1. BioMedLib takes each article matching your query, and summarizes it such that you will spend less time when deciding if the article is useful for you (the screening porcess) and if you should go ahead and obtain the full-text of the article.

The text summarization of BioMedLib has three steps:
1.1 Selecting sentences of the article that contain your query words;
1.2 Selecting sentences that express semantic relation between the query concepts;
1.3 Highlighting important relevant terms, and the relations between them;

And you have the option of viewing the whole abstract (via Pubmed link, like in “view PubMed record for the above article (PMID = 15171183).”), and to customize the highlighting process (change color choices, or turning it off completely).

2. BioMedLib takes all the articles matching your query, and sorts them by their relevance to your question. It will then show the most relevant answers at the top of the results. It saves the users significant time when the best answers are gathered together rather than sprinkled across a large set of results, otherwise the user needs to screen tens or hundreds of articles to find the most relevant ones.

BioMedLib provides three sorting methods:
2.1 By relevance, the superior BML relevance score;
2.2 Sort within a publication date range, and by relevance;
2.3 By publication date;

3. BioMedLib provides link-outs to full-texts of the articles. It will automatically resolve and display content of your subscribed resources. And it clearly marks the articles that have open-access free full text (BML marks such hyperlinks by “[Free full-text]“).

4. BioMedLib provides multiple alternative output options (besides the abstract, and full text).
4.1 Export to citation management software;
4.2 Saving the search results in PDF format;
4.3 Subscription to RSS feeds of your search;
4.4 Sending the search results by email;
4.5 Saving the history of the queries you submitted, and their results.

BioMedLib disambiguation feature

BioMedLib detects unclear phrases in user’s query, and suggests to the user ways to clarify the query interactively, so that to get more precise results.
For example, the query ‘peg complications’ has the word ‘peg’ which can mean different things (hence unclear):
1. Concept C0032478: 400 polyethylene glycol;
2. Concept C0032483: glycol polyethylene;
3. Concept C0072225: 1 2 dihydroxypropane;
4. Concept C0176751: peg percutaneous endoscopic gastrostomy;

Or the query ‘cold’ can mean these different concepts:
1. Concept C0009264: low temperature physical force;
2. Concept C0009443: cold common;
3. Concept C0010412: cold therapy regime therapy;
4. Concept C0024117: cafl chronic airflow limitation;
5. Concept C0234192: cold sensation function;
6. Concept C0719425: cold brand of chlorpheniramine phenylpropanolamine;

BioMedLib provides the suggestions, where each has a checkbox. Then you can simply check the meanings you meant (and uncheck the meanings you don’t want), and resumbit the clarified query. BioMedLib will then automatically screen all the results and choose the ones that match the meaning you intended (the disambiguation process). This saves the user significant time, as the resulting articles are right on target, and the user does not need to screen hundreds of articles searching for the right sense (the intended meaning) of the word in the query.

BioMedLib‘s “related articles”

BioMedLib provides three types of the “related articles” feature.

1. The standard version of the “related articles” feautre, as done in other engines. This focuses on a single article as the starting point, and usually is computed off-line where the results are stored and then shown to user when user asks. This is done in BioMedLib by a query like ’15380493[related]‘ where the integer is the PMID of the starting article and the search tag [related] signifies the user is asking to retrieve related articles to the article with PMID 15380493;

2. BioMedLib enables the user to customize the starting point of the “related articles” with one or a few words.
In the real-world, usually the user starts with a query, which retrieves a set of articles. The user reviews the results and finds one of them to be very relevant. Then the user asks the “related artilces” for that very relevant article. However, in the standard version, there is no way to incorporate the additional important information contained in the original query that user started with. In other words, in the standard “related articles” version, no matter what query the user used to arrive at the relevant article, the results of the “related article” action will be exactly the same.
BioMedLib allows the user to keep the original query as part of the “related articles” computation, thus provding a more precise view of the starting article that matches user’s information needs much better. This is done in BioMedLib by a query like ’15380493[related] heart attack treatment’ where the integer is the PMID of the starting article and the query words ‘heart attack treatment’ express the specific view of the article that the user is interested in.

3. BioMedLib enables the user to start the “related articles” with multiple articles. In other words, the user can ask for “related articles” to a set of given articles, rather than a single article. In BioMedLib a query like ‘(10881502 17437750)[related]‘ will retrieve articles that are related to both PMIDs 10881502 and 17437750. You can choose bigger sets of multiple starting articles (the multi-article relatedness). And you can further customize the query by adding one or a few query words that express your view of interest.

BioMedLib‘s semantic query formulation

If the user is asking for “articles discussing side-effects of meidcations (clinical drugs) in children published in the past 5 years”; how would you do that?
The challenge with queries like the above, is that the user is clearly asking for a concept that contains very many specific instances.
In the above query the user is asking for ‘clinical drugs’, among others. But the user is not specifying a particular instance of drug. In other words any type of “drug” will qualify for the query. Also, any type of the “side-effects” will qualify for the above query. The user is asking for all concepts with the “semantic type” of drug, and all concepts with semantic type of side-effect.

BioMedLib enables the user to search for 135 semantic types, which in turn are organized hierarchically. For example, the query ‘drug[semtyp] injury[semtyp]‘ retrieves all articles that discuss medications and injury or poisoning (side-effects). You can narrow down the query to a range of publication dates like in ‘drug[semtyp] injury[semtyp] 2005:2010[pubdate]‘. And you can use MeSH headings to further narrow down to children, ‘drug[semtyp] injury[semtyp] 2005:2010[pubdate] ((“infant newborn”)[mh] or “infant”[mh] or (“child preschool”)[mh] or “child”[mh] or “adolescent”[mh])’. This will answer the question posed above.

Meaning-based search

BioMedLib provides “meaning-based search” in addition to the widely available “keyword search”.

In a keyword search, the search engine looks for occurrence of user’s words in each document, literally. For example if user submits the query ‘coronary attack‘ the keyword search will return results that are often dramatically different than when user submits the query ‘heart attack‘ .
In the meaning-based search, the search engine is aware that both queries ‘heart attack’ and ‘coronary attack’ point to the same meaning (the same UMLS concept with ID of C0027051). Therefore the BioMedLib search engine will find not only documents that match literally to user’s query words, but also will find all other documents that express the same meaning by using “synonyms” of user’s words. This guaranties the user won’t miss relevant documents just because of the variation in the words the user used to express the question.

Semantic search

BioMedLib provides “semantic search”, on top of the meaning-based search (which in turn is added on top of keyword search).

When you submit a multiword query, of course you like to see documents that contain those words. Besides, the document should have the query words in a “related” fashion.

For example if you submit ‘plastic surgery infection‘ , then not only the documents should contain the three words, but also the documents should explain a sort of relationships between the words. You don’t want a document that talks about infection in one paragraph, and talks about plastic surgery somewhere else and unrelated to the infection; do you?

BioMedLib makes sure documents that express relationship between the concepts of your query, will get higher ranks, so that you find them at the top of the first page of results.

Better keyword search

BioMedLib, besides its meaning-based search, provides a more accurate keyword search as well.

For example submit the double-quoted query “single dose erythromycin” to PubMed, and PubMed will tell you that there is no article in the MEDLINE having the exact phrase “single dose erythromycin”. This is not true; in other words PubMed is missing articles that have the exact phrase “single dose erythromycin”. Try it with BioMedLib , and you will see articles with PubMed IDs 7081971 and 3335066 actually have the exact phrase.

[Kroboth PD, Brown A, Lyon JA, Kroboth FJ, Juhl RP. Pharmacokinetics of single-dose erythromycin in normal and alcoholic liver disease subjects. Antimicrob Agents Chemother. 1982 Jan;21(1):135-40. PMID: 7081971.]

[Malinverni R, Overholser CD, Bille J, Glauser MP. Antibiotic prophylaxis of experimental endocarditis after dental extractions. Circulation. 1988 Jan;77(1):182-7. PMID: 3335066.]

You can try the IDs 7081971 and 3335066 in PubMed and see that the two articles do exist in PubMed’s backend database, but PubMed is unable to find them in response to the exact phrase query “single dose erythromycin” (as of November 2009; and August 2010).

Hassle-free advanced search

The advantages of BioMedLib’s meaning-based search and semantic search come with zero extra hassle for the user. You enter your query as usual, and BioMedLib engine will expand and optimize your query automatically, no questions asked.

And if you don’t like the idea, just uncheck the “expand the query” box in the “[+] Query is expanded” area; and you will get a simple keyword search!
(Note: sometimes the “[+] Query is expanded” area is denoted as “[+] Clarify the query”)

Q1. What is BioMedLib?

BioMedLib™ is a search engine that takes your query and finds the best responses among millions of biomedical articles in the National Library of Medicine’s MEDLINE® database.

Moreover, BioMedLib provides several information services, which are delivered to you via email, free of charge. (see http://bmlsearch.com/?&flnm2=deliveryservices.html)

Q2. What is the advantage of BioMedLib versus other engines?

The differentiating feature of BioMedLib versus other search engines is that BioMedLib finds the most relevant articles for your query and shows them at the top of the list, so that you can locate them much faster, and you can be sure you have located all the relevant articles and missed none.

BioMedLib finds relevant articles that other engines will miss (see an example here http://bmlsearch.com/?&flnm2=pubmedmisses.html). If you want to make sure you are not missing any articles that are relevant to your research, then you should check BioMedLib.

Furthermore, BioMedLib saves you time. BioMedLib sorts the articles so that the most relevant ones show at the top of the list. So you don’t have to spend a significant amount of time screening long lists of articles.

Q3. How is BioMedLib created, and who is behind it?

• Does it take a long time to screen your search results in order to locate relevant articles?
• Are you sure you have found all the relevant publications for your query?
• Do you need to monitor authors who are publishing on your topic?
• Do you wish your search engine could sort the results by their relevance and publication date?
• Do you want to have a PDF copy of the search results for your records?
• Are you tired of using special query syntax language for more relevant results?

Seeking to help themselves and others to overcome their frustrations with the search process, a group of biomedical scientists (we!) used extensive research and their years of experience to build BioMedLib.

The design principles of BioMedLib are taken from the research we have done on information retrieval in health and medicine.

We take two databases, MEDLINE and UMLS, and process them heavily till we end up with a representation that is optimal for BioMedLib. Both the MEDLINE and UMLS are open access, and you can have them for free. However the content they hold are in generic format, which may or may not be suitable for a particular purpose.

The “we” includes me, Mir Siadaty, who creates the differentiating features of BML (read about my education in this blog:  http://mirsiadaty.net/?p=9), plus works of a large number of people, such as multiple independent groups of programmers like www.fedoraproject.org www.apache.org etc., multiple projects in the National Library of Medicine, and multiple processing machines at different levels of intelligence (no kidding! We automate our knowledge-engineering tasks to a very large degree using machines that are intelligent enough to do the tasks correctly, fast, and cheap. We observe that similar organizations/companies spend significant money to hire and train large groups of people to do similar tasks, only slower and more error-prone).

When you have a collection of 10 or 100 documents, you can manually locate any of them quite fast. It is not so fast when the collection grows to thousands.

There are tens of millions of peer-reviewed biomedical publications, and there are multiple search engines to help you locate the document you want.

However, the set of articles the search engines return in response to your query can be large, and it can become very time-consuming to inspect the search results. The BioMedLib.com search engine sorts the retrieved articles such that the most relevant ones are displayed at the top of the first page, saving you time. This is one of the differentiating features of BioMedLib as compared to PubMed.org engine. PubMed can sort the results by publication date but not by relevance.

Authors of biomedical articles use different terms to express the same concept, like in ‘heart attack’ vs ‘myocardial infarction’ vs ‘MI’ vs ‘coronary attack’. The BioMedLib search engine understands 20 million terms expressing 2.3 million unique concepts. BioMedLib automatically incorporates its knowledge in the searches you perform. Therefore no matter which term you choose to formulate your query, and no matter what terms authors have used in their articles, BioMedLib will find the answers. This is one of the differentiating features of BioMedLib as compared to Google.com engine. Google primarily uses a brute-force text-searching and no meaning-based search.

BioMedLib offers several open-access and free-of-charge information services. These are usually delivered to you by e-mail but include real-time online services.

Looking for a Way to Monitor the Biomedical Literature After Your Publication?—Choose “Who is Publishing in My Domain?”

Let’s say you have published an article or you are focused on a specific article for some reason. You would like to know about subsequent publications on the same topic, for defined ranges of time like the 24 months after the publication date.

Register for the “Who is Publishing in My Domain?” service, and you will get periodic e-mails with articles on the same topic as your article. You will be able to choose the frequency of e-mails sent to you and other parameters.

Looking for Articles that Cite Your Article?—Choose “You Are Cited.”

You would like a list of publications (from any domain) that have referenced your article. Then, register for the “You Are Cited” service.

Looking for Recent Articles Related to Your Keywords? —Choose “Article Summarization” or “Updater.”

Now, let’s say that you don’t have a specific article in mind. Instead, you have a few keywords, like the keywords that you use to form a query and submit to a search engine. And let’s say that you like to be informed and up to date about the recent publications that are relevant to the keywords you have. But given your other academic, clinical, research, and administrative responsibilities, you can’t spend much time on scanning multiple articles every month.

The “Article Summarization” service

• locates the recent and relevant articles,
• digests them,
• creates a summary of the most important sentences of the articles, and
• e-mails you the summary statement by the frequency you choose (like every 2 months).

The sentences are hyperlinked to the chosen articles so that you can easily drill down. Reviewing the summary should take a very small amount of your time yet enable you to stay abreast of the most important and recent research findings relevant to your keywords.

On the other hand, if you like to have the list of recent articles relevant to your keywords, without any summarization, then register for the “Updater” service.

Looking for Full-Text Access Even without Journal Subscriptions—Choose “Free PDFs.”

If you like to have access to the full text of the articles, but the computers you use have no subscriptions to journals, which charge to show the full text of the articles, the “Free PDFs” service finds recent publications relevant to your keywords which come with free full text.

Is the “advanced search” of your search engine too complicated to use?

Do you take advantage of the advanced search features of your search engines, so that to formulate better queries and get better results? Do you have to learn and remember copious details about the syntax language used in advanced search? Does the gory detail prevent you to utilize the advanced search?

There are search engines that incorporate advanced search features into every search that you perform, automatically and without any extra effort on your side.

When you submit a query, the search engine automatically expands your query in order to find and distinguish articles where your query words appear in specific locations: article’s title versus abstract versus MeSH terms (major vs minor). The search engine incorporates these additional valuable information (in addition to the fact that the article has your query words ‘somewhere’), so that to better rank the articles and show you the most relevant articles first. For example an article where your query words appear in the title is usually more relevant to your question than when the query words appear in the abstract. When you submit a query like heart attack, the search engine automatically expands the query to something like

(myocardial infarction)[title]^7 or (myocardial infarction)[abstract]^5 or (myocardial infarction)[meshmajor]^6 or (myocardial infarction)[meshminor]^5

The example shows automatic utilization of field-specific searches (all the 15 combinations of the four fields, as well as number of copies per field), and differential importance weights per search field. Also the example implicitly shows that the query has been recognized as a specific biomedical concept, and all the synonyms of the concept are being used in the search, where each synonym has a different Specificity weight.

The BioMedLib search engine, not only expands your query to all the location-specific fields of the article, but also it takes into account co-occurrence of the query words in the same sentence. This is a surrogate for presence of relationship between the query words in the article, sometimes referred to as ‘semantic search’.

DISCLOSURE: Mir implements new algorithms for BioMedLib.com

Do you need to monitor authors who are publishing on your topic?

Inspecting the “body of evidence” is a requirement for scientific authors. As authors of biomedical articles, we are scientifically obligated to do a thorough and unbiased review of the literature when writing our own articles. The advent of modern information retrieval systems has helped the authors in locating relevant prior work. However, the review of literature still is an arduous and time-consuming task.

After an article is published, there is a sharp decline in authors’ review of the literature for new publications on the same topic. Few, if any, dispute the usefulness of monitoring new publications. But the task seems too hard and time consuming for the authors of biomedical publications, given their other clinical, teaching, administrative, and research commitments.

Imagine a service that would update you on newly published material with only minimal effort on your part. You provide

• the ID of your publication (like the PMID of your article, or the citation or title),
• the frequency with which you would like to receive the updates (e.g., every 3 months), and
• your e-mail address.

The rest of the work is done automatically by the service:

• The PMID is used to locate your article,
• then an algorithm is used to rank all the biomedical publications based on their “relatedness” to your article,
• the list of ranked related articles is narrowed down by the publication date (within the previous 3 months), and
• the service e-mails the results to your inbox.

The service automatically repeats the steps based on the frequency that you have requested.

BioMedLib monitors the biomedical literature and updates you on the recent articles published on your topic.

Some search engines can find “related citations” of your article. They usually have the ability to limit a given search to a specific publication date. But they may lack the ability to do both (subsetting by ‘relatedness’ and ‘publication date’) simultaneously, which decreases the usefulness of the results significantly.

The following picture is an example of the content composed and sent by the BioMedLib service.

NOTES:

Since the list of articles are sorted based on their “closeness” to user’s article, then user’s article itself will be at the top of the list. This however does not imply that user’s article is necessarily better than any other article of the list.

DISCLOSURE: Mir implements new algorithms for BioMedLib.com

Do you wish your search engine could sort the results by their relevance and publication date?

Some search engines are able to sort the results by publication date. Sorting by publication date is useful for viewing the recent articles. And recency is an important factor when retrieving, choosing, sorting, and displaying search results.
However, an article that is not relevant to your question is not useful, no matter how recent it is.

Some search engines can sort by relevance (although we need to pay attention to the definition of their relevance score, which is subject of a separate blog post). Sorting by relevance is useful for locating good answers for the question at hand. And relevance is another important factor for retrieval and display of results.
But an article that is too old may not be useful no matter how relevant it is.

Other search engines are capable of both functions (sorting by date and sorting by relevance), but not at the same time.
Articles that are both relevant and recent usually satisfy our information needs much better. Tell me what you think, by leaving a comment.

A method to implement the retrieval of articles which simultaneously satisfy both the ‘recency’ and the ‘relevance’ conditions is for the user to supply her definition of ‘recent’ along with the query words. For example in the query HRT adverse effects 2005:2010[pubdate] the search engine will use the ‘HRT adverse effects’ component of the query to score each article and sort them by their relevance, and use the ’2005:2010[pubdate]‘ component to subset the sorted list of relevant articles to the ones published within the years 2005 to 2010 (which is user’s definition of recency). The result is that the user will get the most relevant article for adverse effects of hormone replacement therapy which are all published recently.

BioMedLib makes it possible to simultaneously sort by relevance and subset by publication date.
After submitting a new query to BioMedLib-which by default sorts the articles by relevance-you will see a few hyperlinks right below the search box:

• “Focus on the recent 5 years”,
• “Focus on the current year”,
• “Focus on the last 30 days”, etc.

Simply click on the desired hyperlink, and BioMedLib will find the articles that are both relevant and recent.
You can find more choices to define what you consider as recent in the ‘Advanced-Search > Publication-Date’ menu.

DISCLOSURE: Mir implements new algorithms for BioMedLib.com

Are you sure you have found all the relevant publications for your query?

Have a look at this example to see how you could miss relevant articles when using other search engines.

• Submit the double-quoted query “single dose erythromycin” to PubMed, and PubMed will tell you that there is no article in the MEDLINE having the exact phrase “single dose erythromycin”. This is not true; in other words PubMed is missing articles that have the exact phrase “single dose erythromycin”.
  Try it with BioMedLib, and you will see articles with PubMed IDs 7081971 and 3335066 actually have the exact phrase.Note: we used PubMed for this example because we are sure you have access to it. Otherwise it is possible to show how other subscription search engines miss relevant articles too.
• Here are the two articles that contain the exact phrase “single dose erythromycin” :
• Kroboth PD, Brown A, Lyon JA, Kroboth FJ, Juhl RP. Pharmacokinetics of single-dose erythromycin in normal and alcoholic liver disease subjects. Antimicrob Agents Chemother. 1982 Jan;21(1):135-40. PMID: 7081971.
• Malinverni R, Overholser CD, Bille J, Glauser MP. Antibiotic prophylaxis of experimental endocarditis after dental extractions. Circulation. 1988 Jan;77(1):182-7. PMID: 3335066.
• You can try the PMIDs 7081971 and 3335066 in PubMed and see that the two articles do exist in PubMed’s backend database, but PubMed is unable to find them in response to the exact phrase query “single dose erythromycin” (as of November 2009; and August 2010).

There are absolute misses in which a search engine is unable to retrieve a relevant article, as in the above example. And then there are practical misses in which a search engine may return the article but displace it at a remote page or bury it within so many irrelevant articles that it becomes practically impossible for the user to find. (The ‘practical misses’ are related to how the search engine sorts the articles in the list of results, and how it displays the results.)

BioMedLib uses advanced algorithms to compute the relevance score for each article, so that all the relevant articles are found and are displayed in the top pages.

DISCLOSURE: Mir implements new algorithms for BioMedLib.com

Does it take a long time to screen your search results in order to locate relevant articles?

You are accustomed to seeing your search results sorted by publication date only. You have to screen ALL of the pages of results, since there is no gradient of relevance across the pages.

For example, for the query ‘breast cancer gene therapy’ PubMed finds over six thousand articles, where the results spread over 300 pages (using the default twenty-article-per-page). Since the results are sorted by publication date, and not by their relevance to your query, the chance that you find a relevant article in page 1 is the same as page 2, or page 10, or page 300. In other words, if you want to make sure you found the relevant articles, you have to screen over 300 pages of results. And that is a long and time-consuming process.

BioMedLib solves this problem by taking your query and calculating probability of relevance for each result. BioMedLib sorts the results so that the articles with the highest probability of relevance show on the first page.

How does sorting-by-relevance make the screening process faster?

Ok, say you are in BioMedLib and you are screening the results page by page. You arrive at a page (it can be page 2, or 10, or 30, depending on your query) where you see that none of the articles on the page are relevant to your query. Then we are sure that there won’t be any relevant articles past that page, so you can stop screening knowing that you have covered all the relevant publications and missed none.
This is a much faster and more reliable screening process (given the fact that we humans are not good at long and monotonous processes).

In addition to the different approach for screening the results mentioned above, a single result page in BioMedLib takes less time to screen than in other search engines. This is the topic of a separate blog post.

Note: you can add some more criteria to the query so that it returns much smaller result set. However, if this is done mainly because it is infeasible to screen so many pages of results, then it has the danger of losing Sensitivity (in other words, increasing the chance of missing relevant articles). The solution provided by BioMedLib to make the screening feasible is more optimal (it has a better Sensitivity + Specificity).

DISCLOSURE: Mir implements new algorithms for BioMedLib.com