Patent Application
20210020313
The present application claims priority from Australian Patent Application No 2018201783 filed on 13 Mar. 2018, the content of which is incorporated herein by reference.
This disclosure relates to generating interactive graphical visualisations of clinical and data.
Clinicians generally examine patients and record their observations (phenotypes). Clinicians also have access to a stock of knowledge from specialists and researchers around the world. However, it is still difficult for clinicians to use this information efficiently. In particular, it is difficult for a clinician to decide which disorders are indicated by the currently observed phenotype.
More particularly, each disease can be defined by a set of phenotypes that are stored in large databases. However, in most cases there is not an exact match between the observed phenotypes and the phenotypes stored for a particular disorder. This makes it difficult for the clinician to explore the disorders that are most relevant for this particular patient.
For example, the Database Online Mendelian Inheritance in Man (OMIM) comprises about 7,500 disorders, which are annotated with phenotypes, where each disorder is associated with about 2-30 phenotypes. For multiple observed phenotypes it therefore quickly becomes impossible to find the most relevant disorders. Even a computer-aided approach would quickly become impractical due to excessive computational complexity and resulting slow response time.
Any discussion of documents, acts, materials, devices, articles or the like which has been included in the present specification is not to be taken as an admission that any or all of these matters form part of the prior art base or were common general knowledge in the field relevant to the present disclosure as it existed before the priority date of each claim of this application.
Throughout this specification the word “comprise”, or variations such as “comprises” or “comprising”, will be understood to imply the inclusion of a stated element, integer or step, or group of elements, integers or steps, but not the exclusion of any other element, integer or step, or group of elements, integers or steps.
There is a need for a computerised tool that the clinician can use and that provides access to the vast amount of data and knowledge that is available. This tool may filter the available options based on the observed phenotype so that the clinician can ultimately find a most relevant disorders.
Disclosed herein is a method that quantifies the similarity between a set of observed phenotypes and a set of stored phenotypes. This set of stored phenotypes may be characterising a disorder or may contain the phenotypes observed on another patient. A quantification of the similarity allows the sorting of candidate diseases (or sets of phenotypes), which allows the reduction of data that is to be provided to a human user. This way, the human user is able to understand the data. For example, the most similar disorders or sets of stored phenotypes may be automatically selected, which allows easy visual inspection of the different associations.
A method for creating a graphical visualisation of clinical data comprises:
receiving the clinical data indicative of multiple observed phenotypes of a patient;
accessing a set of stored phenotypes;
accessing on a database an ontology of phenotypes including hierarchical relationships between the phenotypes of the ontology;
calculating a phenotype-to-phenotype similarity value indicative of a similarity between each of the observed phenotypes and each phenotype in the set of stored phenotypes, based on the ontology;
determining an assignment of one stored phenotype of the set to each of the observed phenotypes based on the phenotype-to-phenotype similarity values;
aggregating the phenotype-to-phenotype similarity values of the stored phenotypes from the set that are assigned to each of the multiple observed phenotypes into a set-to-set similarity value indicative of a similarity between the observed phenotypes and the set of stored phenotypes;
repeating the accessing, calculating, determining the assignment and aggregating steps for each of the multiple sets of stored phenotypes to thereby calculate a set-to-set similarity measure for each of the multiple sets;
selecting one or more of the multiple sets based on the aggregated set-to-set similarity values; and
generating a graphical user interface comprising a graphical indication of the selected one or more of the multiple sets in relation to the multiple observed phenotypes.
It is an advantage that the similarity between the observed phenotypes and the sets of phenotypes is determined based on the distance in the ontology. This way, inexact matches can be considered and the candidate diseases can be selected for the clinician. Further, determining an assignment enables the use of computationally efficient heuristic algorithms which reduce the time required for computation. Together with the use of an ontology this allows rapid calculations leading to an enhanced user experience. For example, the clinician can select different patients and different disorders and immediately receive a selection of most relevant candidates without having to wait for complex calculations to be completed.
Determining the assignment may comprise determining an assignment by optimising a cost that is based on the phenotype-to-phenotype similarity values.
Determining the assignment may comprise applying a heuristic to determine the assignment by selecting one assignment at a time with optimal cost and then determining remaining assignments.
Determining the assignment may comprise performing an Hungarian algorithm.
Aggregating the phenotype-to-phenotype similarity values may comprise calculating an average of the phenotype-to-phenotype similarity values.
The method may further comprise splitting observed phenotypes and stored phenotypes by anatomical systems and aggregating set-to-set similarity values across the anatomical systems.
Aggregating across the anatomical systems may comprise calculating an average of the set-to-set similarity values the anatomical systems.
Generating the user interface may comprise generating a graphical indication of the phenotype-to-phenotype similarity values.
The graphical indication of the phenotype-to-phenotype similarity value may comprise a line with a first visual appearance for an exact match and a second visual appearance for an inexact match.
The set of stored phenotypes may be associated with a disorder.
The set of stored phenotypes may be associated with a further patient.
Calculating the phenotype-to-phenotype similarity value may comprise determining a distance in the ontology from the observed phenotype to each phenotype in the set of stored phenotypes.
Calculating the phenotype-to-phenotype similarity value may be based on an information content of the observed phenotype in the ontology and an information content of the stored phenotype in the ontology and an information content of a least common subsumer of the observed phenotype and the stored phenotype in the ontology.
The information content may be based on a count of leaf nodes under children of the phenotype in the ontology, a count of ancestors of the phenotype in the ontology and the total number of leaf nodes in the ontology.
A computer system for creating a graphical visualisation of clinical data comprises:
a data port to receive the clinical data indicative of multiple observed phenotypes of a patient;
a data store from which to access a set of stored phenotypes;
database to store an ontology of phenotypes including hierarchical relationships between the phenotypes of the ontology;
a processor to:
Optional features described of any aspect of method, computer readable medium or computer system, where appropriate, similarly apply to the other aspects also described here.
An example will be described with reference to:
The computer system 100 comprises a processor 104 connected to program memory 105, data memory 106, a communication port 107 and a database 108. When reference is made herein to a database, it is to be understood as any form of structured data storage including comma separated values, SQL or graph based databases, which are preferred due to their inherent ability to efficiently store and retrieve graph data as used herein.
The program memory 105 is a non-transitory computer readable medium, such as a hard drive, a solid state disk or CD-ROM. Software, that is, an executable program stored on program memory 105 causes the processor 104 to perform the method in
The processor 104 may then store the graphical user interface on data store 106, such as on RAM or a processor register. Processor 104 may also send the graphical user interface via communication port 107 to client device 102 such as through the use of a web server installed on computer system 100 and a browser application installed on client device 102.
The processor 104 may receive data, such as clinical data, from data memory 106 as well as from the communications port 107. In one example, the processor 104 receives clinical data from client device 102 via communications port 107, such as by using a Wi-Fi network according to IEEE 802.11. The Wi-Fi network may be a decentralised ad-hoc network, such that no dedicated management infrastructure, such as a router, is required or a centralised network with a router or access point managing the network.
In one example, processor 104 receives and processes the clinical data in real time. This means that the processor 104 creates the graphical user interface every time clinical data is received from client device 102 and completes this step before the client device 102 sends the next clinical data update. The same may apply for re-arranging the graphical user interface such that the time between the user interacting with the graphical user interface and the graphical user interface being updated on client device 102 is not perceived as a delay, such as less than 1 s or less than 100 ms. User interaction may comprise selection of sets of stored phenotypes, such as sets associated with further patients or sets associated with disorders of interest.
Although communications port 107 is shown as distinct module, it is to be understood that any kind of data port may be used to receive data, such as a network connection, a memory interface, a pin of the chip package of processor 104, or logical ports, such as IP sockets or parameters of functions stored on program memory 104 and executed by processor 104. These parameters may be stored on data memory 106 and may be handled by-value or by-reference, that is, as a pointer, in the source code.
The processor 104 may receive data through all these interfaces, which includes memory access of volatile memory, such as cache or RAM, or non-volatile memory, such as an optical disk drive, hard disk drive, storage server or cloud storage. The computer system 100 may further be implemented within a cloud computing environment, such as a managed group of interconnected servers hosting a dynamic number of virtual machines.
It is to be understood that any receiving step may be preceded by the processor 104 determining or computing the data that is later received. For example, the processor 104 determines clinical data and stores the clinical data in data memory 106, such as RAM or a processor register. The processor 104 then requests the data from the data memory 106, such as by providing a read signal together with a memory address. The data memory 106 provides the data as a voltage signal on a physical bit line and the processor 104 receives the clinical data via a memory interface.
It is to be understood that throughout this disclosure unless stated otherwise, nodes, edges, graphs, solutions, variables, paths, sets and the like refer to data structures, which are physically stored on data memory 106 or processed by processor 104. Further, for the sake of brevity when reference is made to particular variable names, such as “similarity value” or “distance” this is to be understood to refer to values of variables stored as physical data in computer system 100.
It is noted that for most humans performing the method 200 manually, that is, without the help of a computer, would be practically impossible. Therefore, the use of a computer is part of the substance of the invention and allows using the available data that would otherwise not be possible or prohibitively difficult due to the large amount of data and the large number of calculations that are involved.
In order to address this issue, processor 104 accesses 203 on database 108 an ontology of phenotypes including hierarchical relationships between the phenotypes of the ontology. While database 108 is shown as integral part of computer system 108, it may equally be hosted externally, such as on a publicly available cloud computing environment. In one example, clinician 101 enters observations as text in natural language and a natural language processor analyses the text input and maps it to a phenotype ontology, such as phenotypes included in OMIM or to the Human Phenotype Ontology (http://human-phenotype-ontology.github.io, HPO). As described on their website the HPO is a computational representation of a domain of knowledge based upon a controlled, standardized vocabulary for describing entities and the semantic relationships between them.
The HPO aims to provide a standardized vocabulary of phenotypic abnormalities encountered in human disease. Each term in the HPO describes a phenotypic abnormality, such as atrial septal defect. The HPO is currently being developed using the medical literature, Orphanet, DECIPHER, and OMIM. HPO currently contains approximately 11,000 terms (still growing) and over 115,000 annotations to hereditary diseases. The HPO also provides a large set of HPO annotations to approximately 4000 common diseases.
Processor 104 calculates 204 a phenotype-to-phenotype similarity value indicative of a similarity between the observed phenotype 304 and a stored phenotype 306. Processor 104 performs this calculation by determining a distance in the ontology 300 from the observed phenotype 304 to the stored phenotype 306. The distance from observed phenotype 304 to stored phenotype can be computed as the distance from the root 301 to the observed phenotype 304, plus the distance from the root to the stored phenotype 306, minus twice the distance from the root to their lowest common ancestor, which would be node 303 in this case. Therefore, the distance in this case would be 2+3-2*1=3. Further details can be found in: Djidjev H. N., Pantziou G. E., Zaroliagis C. D. (1991) Computing shortest paths and distances in planar graphs. In: Albert J. L., Monien B., Artalejo M. R. (eds) Automata, Languages and Programming. ICALP 1991. Lecture Notes in Computer Science, vol 510. Springer, Berlin, Heidelberg, which is incorporated herein by reference. In another example, the similarity value is computed by
where c1 and c2 are ontological concepts, lcs is the least common subsumer of c1 and c2 and ic is the information content of c as defined in Lin, D.: An Information-Theoretic Definition of Similarity. In: Proc. of Conf. on Machine Learning, pp. 296-304 (1998), which is incorporated herein by reference.
While Lin uses the Resnik model to compute the ic it may be preferable to instead use:
where cleaves is the count of the leaf nodes under all children of c, cancestors is the count of all ancestors of c, maxleaves is the total number of leaf nodes in the ontology. Further information can be found in Seco, N., Veale, T., Hayes, J. An Intrinsic Information Content Metric for Semantic Similarity in WordNet. Proceedings of the 16th European Conference on Artificial Intelligence, ECAI'2004 noting that Seco uses maxnodes instead of maxleaves in their formula—i.e., the total count of nodes in the ontology.
Processor 104 repeats this calculation for each combination of observed phenotype with stored phenotype in the particular set so as to calculate a phenotype-to-phenotype similarity value indicative of a similarity between each of the observed phenotypes and each phenotype in the set of stored phenotypes, by determining a distance in the ontology from the observed phenotype to each phenotype in the set of stored phenotypes. For example, processor 107 may loop over all disorders in the database and for each disorder retrieve the set of phenotypes that define that disorder. Processor 104 may then perform a first loop over all stored phenotypes in that set and perform a second inner loop over the observed phenotypes and calculate the similarity value within the three nested loops (disorders, stored phenotypes and observed phenotypes).
Since this calculation can be relatively complex due to the large number of inner loops (combinations) the computation time can be reduced by splitting the phenotypes into the different anatomical systems, such that processor 104 never attempts to calculate a similarity value between phenotypes from different systems. For example, if there are 4,000 common diseases in the database with each having on average 8 phenotypes, there are 32,000 iterations in the first two loops. For 10 observed phenotypes this would result in 320,000 iterations in the innermost loop. Assuming 1,000 similarity measures can be determined per second, this would lead to 320 seconds (5 minutes) which is too long for a response user interface. Splitting the phenotypes into about 10 anatomical systems, for example, would mean that a large number of combination would not need to be calculated which would reduce the number of inner iterations in some examples by a factor of 10 to about 32 seconds which is more suitable for an entire rebuild of the disease database from scratch. It is a further advantage that the split along the top-level abnormalities (or anatomical systems) also keeps phenotypes localised—i.e., there would otherwise be a similarity value between large head (skeletal) and cafe-au-laix spots (skin), which, from a medical perspective, is not practical.
Once the phenotype-to-phenotype similarity values are calculated, processor 104 determines 205 an assignment of one stored phenotype of the set to each of the observed phenotypes based on the phenotype-to-phenotype similarity values. Fields with bold outlines indicate the assignment of a stored phenotype to an observed phenotype. As can be seen in
In one example, processor 104 performs the Hungarian algorithm described in Kuhn, H. W. (1955), The Hungarian method for the assignment problem. Naval Research Logistics, 2: 83-97, which is included herein by reference. The Hungarian algorithm works by first expanding the matrix to a square matrix, finding the minimum cost in each row and subtracting that cost from that row so as to generate one or more zero values. The same is then done for the rows. Processor 104 then determines a selection of zero values to cover the entire matrix by the minimum number of lines (rows or columns). If the number of selected rows/columns is less than the number of rows/columns of the matrix, processor 104 repeats the process. In this sense, processor 104 applies a heuristic to determine the assignment by selecting one assignment at a time with optimal cost and then determining remaining assignments. In one example, processor 104 executes code from the munkres Python module or the scipy.optimize.linear_sum_assignment Python module to determine the assignment.
Next, processor 104 aggregates 206 the phenotype-to-phenotype similarity values 500 of the stored phenotypes from the set that are assigned to each of the multiple observed phenotypes into a set-to-set similarity value indicative of a similarity between the observed phenotypes and the set of stored phenotypes. In this example, this aggregation comprises the calculation of an average value 501, which is ‘2.14’ in this example.
As mentioned above, processor 104 may split the observed phenotypes and stored phenotypes by anatomical systems and aggregating set-to-set similarity values across the anatomical systems. For example, P1, P2 and P3 may relate to the skeletal system, whereas P4, P5, P6 and P7 relate to the digestive system. In this case, the result of the assignment would be the same as before but the calculation to determines the assignment would be significantly reduced because the number of phenotypes in each set is reduced. In the example of split phenotypes, processor would calculate one average per system, that is, (0+2+5)/3=2.33 and (3+3+2+0)/4=2. Processor 104 can then aggregate the two results to calculate (2.33+2)/2=2.17. As can be seen, the difference between two methods is not significant but the reduction in computation time is significant.
While the above examples calculate averages, other aggregation methods may be used, such as sums, squared sums, etc. For example, processor 104 may simply sum up the cost values for the different systems into one sum and then divide by the number of phenotypes.
Processor 104 then repeats 207 the accessing 203, calculating 204, determining the assignment 205 and aggregating 206 steps for each of the multiple sets of stored phenotypes to thereby calculate a set-to-set similarity measure for each of the multiple sets. In other words, the processor 104 keeps the observed phenotypes for each iteration and calculates a set-to-set similarity between the set of observed phenotypes and each set of stored phenotypes, such as phenotypes defining disorders or being associated with other patients.
Once the set-to-set similarity values are calculated, processor 104 selects 208 one or more of the multiple sets based on the aggregated set-to-set similarity value. For example, processor 104 selects the highest ranked sets, such as the top 10 or top 4 sets or all sets that are above a threshold. This way, the number of sets (i.e. disorders) can be reduced from thousands to less than ten or less than five.
Processor 104 then generates 209 a graphical user interface comprising a graphical indication of the selected one or more of the multiple sets in relation to the multiple observed phenotypes. This may involve generating a user interface on a screen directly connected to computer system 100 where the processor 104 performs the calculations. It may also involve generating the user interface in the form of web-accessible content, such as HTML and JavaScript. Client 102 can then access the web-accessible content and render the graphical user interface on a screen of client device 102. Various different front-end/back-end platforms may be used including an Angular/Flask framework.
User interface 600 also includes a graphical indication of the phenotype-to-phenotype similarities between the phenotypes in the selected sets 601, 602, 603 and the observed phenotypes 604. For example, processor 104 may generate a line between phenotypes that are similar. More particularly, processor 104 may generate a solid line between phenotypes that are an exact match (zero distance in the ontology graph) and dashed lines for inexact matches. There may be a threshold on the distance, such as 10, above which processor 104 draws no line.
Clinician 101 can now very clearly see which disorders are similar to the observed set of phenotypes and can also see which phenotypes are similar to understand the determined similarity. This means the method provides clinician 104 with guidance without taking control from the clinician's hands and without withholding or hiding important information from the clinician. In other words, the individual phenotypes are all displayed so that clinician 101 can make a professional conclusion but the data that is irrelevant is filtered out so as to provide a clear view on the data that is relevant.
While the above explanation and in particular
It will be appreciated by persons skilled in the art that numerous variations and/or modifications may be made to the above-described embodiments, without departing from the broad general scope of the present disclosure. The present embodiments are, therefore, to be considered in all respects as illustrative and not restrictive.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2018201783 | Mar 2018 | AU | national |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/AU2019/050221 | 3/12/2019 | WO | 00 |
