25th Australasian Joint Conference on Artificial Intelligence (AI 2012) Editors : Sankalp Khanna, Abdul Sattar, David Hansen
2012 held in conjunction with the
WORKSHOP PROCEEDINGS
Sankalp Khanna (CSIRO Australian e-Health Research Centre, Australia) Aditya Ghose (University of Newcastle, Australia) Anthony Maeder (University of Western Sydney, Australia) Wayne Wobcke (University of New South Wales, Australia) Mehmet Orgun (Macquarie University, Australia) Yogesan (Yogi) Kanagasingam (CSIRO Australian e-Health Research Centre, Australia)
Simon McBride (CSIRO Australian e-Health Research Centre) Adam Dunn (University of New South Wales) Stephen Anthony (University of New South Wales) Lawrence Cavedon (Royal Melbourne Institute of Technology / NICTA) Diego Mollá Aliod (Macquarie University) Michael Lawley (CSIRO Australian e-Health Research Centre) Anthony Nguyen (CSIRO Australian e-Health Research Centre) Amol Wagholikar (CSIRO Australian e-Health Research Centre) Bevan Koopman (CSIRO Australian e-Health Research Centre) Kewen Wang (Griffith University) Vladimir Estivill-Castro (Griffith University) John Thornton (Griffith University) Bela Stantic (Griffith University) Byeong-Ho Kang (University of Tasmania) Justin Boyle (CSIRO Australian e-Health Research Centre) Guido Zuccon (CSIRO Australian e-Health Research Centre) Hugo Leroux(CSIRO Australian e-Health Research Centre) Alejandro Metke (CSIRO Australian e-Health Research Centre) CSIRO Australian e-Health Research Centre Institute for Integrated and Intelligent Systems, Griffith University 9:00 am – 10:30 am 11:00 am – 12:30 pm 2:00 pm – 3:30 pm 3:30 pm – 4:00 pm 4:00 pm – 5:30 pm 5:30 pm
8:30 am – 9:00 am
10:30 am – 11:00 am
12:30 pm – 2:00 pm
PROGRAM
Dr. Jia-Yee Lee National ICT Australia (NICTA)
Dr. Sankalp Khanna CSIRO Australian e-Health Research Centre
Colin Curtain, Ivan Bindoff, Juanita Westbury and Gregory Peterson
Mani Abedini, Michael Kirley and Raymond Chiong
Zarita Zainuddin, Pauline Ong and Kee Huong Lai
Amol Wagholikar, Guido Zuccon, Anthony Nguyen, Kevin Chu, Shane Martin, Kim Lai and Jaimi Greenslade
Zahra Shahabi Kargar, Sankalp Khanna and Abdul Sattar
Qing Zhang, Yang Xie, Pengjie Ye and Chaoyi Pang
Shima Ghassem Pour, Peter Mc Leod, Brijesh Verma and Anthony Maeder
Luke Butt, Guido Zuccon, Anthony Nguyen, Anton Bergheim and Narelle Grayson
Prajni Sadananda and Sadananda Ramakoti
Dr. Christian Guttman IBM Research Australia
Panel Chair : Professor Abdul Sattar . Panelists : Dr. Jia-Yee Lee, Dr. Christian Guttman, Prof. Wayne Wobcke, Prof. Sadananda Ramakoti
5 7 9 11 21 33 45 55 65 77 83 89
Sankalp Khanna1, 2, Abdul Sattar2, David Hansen1 1The Australian e-Health Research Centre, RBWH, Herston, Australia {Sankalp.Khanna, David.Hansen}@csiro.au 2 Institute for Integrated and Intelligent Systems, Griffith University, Australia
A.Sattar@griffith.edu.au 1
The business of health service delivery is a complex one. Employing over 850,000 people, and delivering services to 21.3 million residents, the Australian healthcare system is currently struggling to deal with increasing demand for services, and an acute shortage of skilled professionals. The National eHealth Strategy drives a nationwide agenda to provide the infrastructure and tools required to support the planning, management and delivery of health care services. National initiatives such as the National Health Reform Program, the National Broadband Network, and the Personally Controlled Electronic Health Record are accelerating the use of information and communication technologies in delivering healthcare services. The Australasian Joint Conferences on Artificial Intelligence (AI) provide an excellent opportunity to bring together artificial intelligence researchers who are working in health research.
Driven by a senior program committee comprising distinguished faculty from several Australian universities including Griffith University, the University of New South Wales, University of Newcastle, University of Western Sydney, and Macquarie University, and specialist health research organisations including the CSIRO Australian e-Health Research Centre, the Artificial Intelligence in Health workshop series was created in 2011 to bring these researchers together as part of Australia’s premier Artificial Intelligence conference.
Held for the first time in December 2011, the workshop was the first of its kind to bring together scholars and practitioners nationally in the field of Artificial Intelligence driven Health Informatics to present and discuss their research, share their knowledge and experiences, define key research challenges and explore possible collaborations to advance e-Health development nationally and internationally. The workshop was co-located with the 24th Australasian Joint Conference on Artificial Intelligence and was attended by 25 delegates.
Of the 16 submissions received, 6 were accepted as Full Papers and 5 as Short Papers accompanied with posters. All papers presented at the AIH 2011 workshop were also invited to revise and submit their manuscripts for inclusion in a special issue of the Australasian Medical Journal. Of these, seven papers and a letter to the editor were published in the special issue in September 2012. 3
The Second Australian Workshop on Artificial Intelligence (AIH 2012) is being held in conjunction with the 25th Australasian Joint Conference on Artificial Intelligence (AI 2012) in Sydney, Australia, on the 4th of December, 2012. The Call for Papers received an excellent response this year. All submitted papers went through a rigorous review process. Of these, 6 full papers and 3 short papers have been accepted for presentation in the workshop and for publication in these CEUR proceedings. The workshop will also feature three keynote addresses and a panel discussion on the topic “AI in Health: the 3 Big Challenges”.
This year again, the workshop is offering 4 travel scholarships of $250 each to students who were first authors of accepted papers. A best paper prize of $250 will also be awarded on the workshop day. Both prizes have been sponsored by the CSIRO Australian e-Health Research Centre. All accepted full papers and short papers will be also invited to extend and reformat their papers for publication in a special issue of the Australasian Medical Journal (www.amj.net.au). The journal is indexed on the following databases: DOAJ, EBSCO, Genamics journalseek, ProQuest, Index Copernicus, Open J-Gate, Intute, Global health and CAB Abstracts databases, MedWorm, Scopus, Socolar, PMC, PubMed. 4 4.1
4.2
Sankalp Khanna (CSIRO Australian e-Health Research Centre, Australia) 4.3
Aditya Ghose (University of Newcastle, Australia) Anthony Maeder (University of Western Sydney, Australia) Wayne Wobcke (University of New South Wales, Australia) Mehmet Orgun (Macquarie University, Australia) Yogesan (Yogi) Kanagasingam (CSIRO Australian e-Health Research Centre, Australia) 4.4
Simon McBride (CSIRO Australian e-Health Research Centre) Adam Dunn (University of New South Wales) Stephen Anthony (University of New South Wales) Lawrence Cavedon (Royal Melbourne Institute of Technology / NICTA) Diego Mollá Aliod (Macquarie University) Michael Lawley (CSIRO Australian e-Health Research Centre) Anthony Nguyen (CSIRO Australian e-Health Research Centre) Amol Wagholikar (CSIRO Australian e-Health Research Centre) Bevan Koopman (CSIRO Australian e-Health Research Centre) Kewen Wang (Griffith University) Vladimir Estivill-Castro (Griffith University) John Thornton (Griffith University) Bela Stantic (Griffith University)
Byeong-Ho Kang (University of Tasmania) Justin Boyle (CSIRO Australian e-Health Research Centre) Guido Zuccon (CSIRO Australian e-Health Research Centre) Hugo Leroux(CSIRO Australian e-Health Research Centre) Alejandro Metke (CSIRO Australian e-Health Research Centre) 4.5
CSIRO Australian e-Health Research Centre Institute for Integrated and Intelligent Systems, Griffith University 4.6
The Australasian College of Health Informatics The Australasian Medical Journal The Australasian Telehealth Society 5
We are especially thankful to the organising committee of the 25th Australasian Joint Conference on Artificial Intelligence (AI 2012). This workshop series would not have possible without their support. We would also like to thank the Workshop Chair of AI 2012, Hans Guesgen, for organising the workshops and championing these CEUR workshop proceedings.
Dr. Jia-Yee Lee National Information and Communications Technology Australia Ltd (NICTA), Australia jia-yee.lee@nicta.com.au
Dr Jia-Yee Lee is the Director of the Health and Life Science Business Team at National ICT Australia Ltd. She manages the business, commercial and research activities of the NICTA groups in Diagnostic and Computational Genomics, Biomedical Informatics, Portable Motion Analytics and Bio-Imaging Analytics. Prior to joining NICTA, Jia-Yee spent 10 years in the management consulting sector providing leadership in developing and implementing strategies and operational plans that improved business outcomes for clients in government, ICT, and healthcare sectors. Her business plans have led to international and national investments into Australian-based start-ups. Jia-Yee has extensive experience as a project manager working on complex multi-disciplinary and multi-million dollar programs funded by State and Commonwealth governments. Her ehealth experience includes stakeholder engagement with clinicians and leading technical teams to implement a range of commercial web-based systems for the healthcare and medical research sectors. With more than 20 years in medical research, Jia-Yee has led programs at MacFarlane Burnet Centre (now "Burnet Institute") and the Victorian Infectious Diseases Reference Laboratory, Melbourne Health. Her research into hepatitis B virus and rubella virus was funded by the National Health and Medical Research Council of Australia. Jia-Yee’s research skills include molecular and diagnostic virology, and electron and confocal microscopy. Jia-Yee has a PhD from the University of Melbourne and a MBA from Melbourne Business School.
Sankalp Khanna The Australian e-Health Research Centre, RBWH, Herston, Australia
Sankalp.Khanna@csiro.au
Sankalp is a Postdoctoral Fellow at the Australian e-Health Research Centre, the leading national research facility applying information and communication technology to improve health services and clinical treatment for Australians. As a member of the Forecasting and Scheduling team, he is actively engaged in projects in the areas of planning and optimization, patient flow analytics, prediction and forecasting, and predictive scheduling, all aimed at employing artificial intelligence to improve the efficiency of the health system.
His research interests include Applied Artificial Intelligence, Prediction and Forecasting, Planning and Scheduling, Multi Agent Systems, Distributed Constraint Reasoning, and Decision Making and Learning under Uncertainty.
Sankalp completed a PhD in 2010 looking at intelligent techniques to
model and optimise the complex, dynamic and distributed processes of
Elective Surgery Scheduling. He was the recipient of a state award for
outstanding
Sankalp was also the founding workshop chair of this AI in Health workshop series.
Christian Guttman
IBM Research, Australia
Christian.guttmann@au1.ibm.com
Dr. Guttmann leads and defines projects around health care at the newly established IBM Research labs in Melbourne – the 11th lab of IBM Research worldwide. One focus of Guttmann’s work is to build smarter analytics that enables health care entities (doctors, nurses, hospitals, pharmacies, etc) to collaborate more efficiently in complex environments. His work addresses the information and communication challenges faced by tomorrow’s world of health care: How can we create and apply smarter collaborative health care technologies that cope with the tsunami of chronic diseases.
Prior to IBM, Dr. Guttmann led the research theme on health care and disaster at the Etisalat British Telecom Innovation Centre (EBTIC). The theme partnered with major stakeholders, including governmental health authorities and ministries. He has been a research fellow at the Faculty of Medicine, Nursing and Health Sciences at Monash University, where he researched how intelligent systems can improve collaborative care (done in together with primary health care providers). He worked also in industrial projects with HP and Ericsson.
Dr. Guttmann holds a PhD degree from Monash University, two Master degrees from Paderborn University (Germany) and the Royal Institute of Technology (Sweden), and a psychology degree from Stockholm University (Sweden). He organised major conferences and workshops, edited two books on intelligent agent technologies, and co-authored over 30 articles in leading conferences and journals.
An investigation into the types of drug related problems that can and cannot be identified by commercial medication review software Colin Curtain, Ivan Bindoff, Juanita Westbury and Gregory Peterson
Unit for Medication Outcomes Research and Education
School of Pharmacy
University of Tasmania {Colin.Curtain, Ivan.Bindoff, Juanita.Westbury, G.Peterson}@utas.edu.au Abstract.
A commercially used expert system using multiple-classification rippledown rules applied to the domain of pharmacist-conducted home medicines review was examined. The system was capable of detecting a wide range of potential drug-related problems. The system identified the same problems as pharmacists in many of the cases. Problems identified by pharmacists but not by the system may be related to missing information or information outside the domain model. Problems identified by the system but not by pharmacists may be associated with system consistency and perhaps human oversight or human selective prioritization. Problems identified by the system were considered relevant even though the system identified a larger number of problems than human counterparts.
Keywords: Clinical decision support system, multiple-classification rippledown rules, expert system, pharmacy practice 1
Introduction
A drug-related problem (DRP) can be broadly defined as “…an event or circumstance involving drug therapy that actually or potentially interferes with desired health outcomes”[1] DRPs comprise a spectrum of problems including over- or underdosage, drug-drug or drug-disease interactions, untreated disease and drug toxicity. Patient health education and compliance with therapy may be sub-standard and subsequently also be considered as drug-related problems. DRPs can be dangerous; For instance, a marginally high daily dose of warfarin has the potential to cause fatal bleeding.
Home medicines review (HMR) is a Commonwealth Government funded service conducted by accredited pharmacists to identify and address DRPs among eligible patients [2]. The main aims of the service are to enhance patient knowledge, quality use of medicines, reconcile health professional awareness of actual medication use and, ultimately, improve patient quality of life. The HMR service is a collaborative activity between health professionals, typically accredited pharmacists, general practitioners (GPs), and patients. Since its inception in 2001 the service has steadily grown with nearly 80,000 HMRs funded in the 2011/2012 period [3].
An HMR is initiated for eligible consenting patients by a GP. Eligible patients are identified if they regularly take 5 or more medications among other criteria [2]. An HMR accredited pharmacist then obtains medical information from the GP, covering medical history, current medications and pathology.
A core component of an HMR is an interview between the pharmacist and the patient, with interview typically conducted in the patient’s home. The interview, elicits additional information such as: actual medication use, additional non-prescribed medications, an understanding of the patient’s motivation behind actual rather than directed medication use, and the patient’s health and medication knowledge [4]. This process allows for a deeper understanding of the patient’s situation and gives the pharmacist insight into cultural or language barriers, physical and economic limitations and family support.
The amassed information is reviewed by the pharmacist to identify actual and potential DRPs. The pharmacist writes a report of findings for the patient’s GP, which includes recommendations to resolve any actual or potential problems. Consultation between the GP and the patient culminates in an actionable medication management plan designed to trial changes to existing therapy, and ideally, lead to improved medication use and improved patient health outcomes [4].
An important component is the professional skill of the pharmacist to be able to identify clinically relevant DRPs from the available information. This requires a wide scope of knowledge, not only of medications, but of evidence-based guidelines and contemporary management of a variety of medical conditions.
Evidence-based guidelines can be difficult to implement due to their apparent complexity. An example is provided from Basger et al.’s Prescribing Indicators in Elderly Australians: “Patient at high risk of a cardiovascular event (b) is taking an HMG-CoA reductase inhibitor (statin)”[5] If a patient did not meet this criterion this would be considered a DRP. It can be reasonably expected that pharmacists would be aware of statin medications currently available in Australia, in October 2012 these were: atorvastatin, fluvastatin, pravastatin, rosuvastatin, and simvastatin. Note (b) specifies those patients at high risk of cardiovascular event: “age>75 years, symptomatic cardiovascular disease (angina, MI[myocardial infarction], previous coronary revascularization procedure, heart failure, stroke, TIA[transient ischemic attack], PVD[peripheral vascular disease], genetic lipid disorder, diabetes and evidence of renal disease (microalbuminuria and/or proteinuria and/or GFR[glomerular filtration rate]<60ml/min”. Determining patients at high risk of cardiovascular events is more problematic and requires sufficient additional information to make such a determination. One obvious problem is the amount of information that needs to be screened, both within the guideline text and the patient data, to identify appropriate patients.
A commercial product developed by Medscope, Medication Review Mentor (MRM)[6], incorporates a clinical decision support (CDSS) tool to assist with the detection of DRPs. MRM utilizes a knowledge-based system to detect DRPs and provide recommendations for their resolution. This knowledge-based system uses the multiple classification ripple-down rules (MCRDR) method and was based on the work of Bindoff et al. who applied this approach to the knowledge domain of medication reviews [7, 8]. The ripple-down rules method was considered appropriate as knowledge could be gradually added to the knowledge base, broadening the scope and refining existing knowledge as the system was being used [7, 9]. Bindoff et al. suggested intelligent decision support software developed for this knowledge domain may improve the quality and consistency of medication reviews.
No prior research had been undertaken to determine the clinical decision support capacity of this commercial software, apart from contemporary research by the authors. This contemporary research by the authors assessed opinions from pharmacology experts and had determined that MRM is capable of identifying clinically relevant DRPs [10-12].
This evaluation attempts to provide light on the scope of DRPs that can be identified by this software by presenting summary counts and examples of the types of problems that were identified by MRM and by pharmacists. This paper evaluates the similarities and differences between pharmacist findings and MRM findings more in terms of a qualitative comparison by highlighting common findings, extremes of difference and discussing the possible advantages and limitations of the software, as well as discussing areas for potential improvements. 2
How MRM works
The decision support component of MRM is a knowledge-based system which uses MCRDR as its inference engine. MCRDR provides the knowledge engineer a way to incrementally improve the quality of the knowledge base through the addition of either new rules – which are added when the system fails to identify a DRP, or refinements to existing rules – which are added when the system incorrectly identifies an inappropriate DRP. The system’s knowledge base is managed by medication review experts, who regularly review cases, examining the findings of the system for that case, and then adding/refining rules until the system produces a wholly correct set of findings for that case [8]. The validity of new rules is always being ensured, as the system identifies any conflicts which may arise from the addition of the new rule, and prompts the pharmacist to refine their rule until no further conflicts arise. 3
Methods
Australia-wide data collected during 2008 for a previous project, examining the economic value of HMRs, was used for this study [13]. The data contained patient demographics, medications, diagnoses and pathology results for 570 communitydwelling patients aged 65 years old and older. The 570 HMRs were obtained from 148 different pharmacists. Supplementing this data were the original reviewing pharmacists’ findings, detailing pharmacist-identified DRPs and recommendations.
The HMR data were entered into MRM and DRPs identified by MRM were recorded. MRM utilized a wide range of information including basic patient demographics such as age and gender, medication type including strength, directions and daily dose. MRM could calculate daily dose from strength and directions in many cases. Duration of use of medication could be entered, which included options of less than 3 months and more than 12 months. Medications were assigned Anatomic Therapeutic Chemical classifications (ATC) [14]. ATC is a five-tier hierarchical classification system allowing medications with similar properties to be grouped together in chemical classes which are then grouped into therapeutic categories.
Diagnoses could be entered and were based on the ICPC2 classifications [15]. The ICPC2 classification system was also hierarchical, grouping diagnoses under similar categories. Diagnoses could be assigned temporal context as recent, ongoing or past history. Medication allergies and general observations including height, weight and blood pressure could be entered. A wide range of pathology readings could be entered, including biochemical and hematological data.
At the time of the data entry and collections of results, August 2011, MRM contained approximately 1800 rules [16]. Rule development was undertaken by a pharmacist with expertise in both clinical pharmacology and HMRs [6].
Direct comparison of the DRPs identified by MRM and those identified by the original pharmacists was not possible due to the individual textual nature of each DRP. Each DRP identified by either the pharmacist or MRM was mapped to a concept (defined here as a theme) that described the DRP in sufficient detail to allow comparisons of similarity and difference between pharmacists and MRM. The themes often described the type of drug or disease and other relevant factors involved. The development of a list of themes and the mapping of DRPs to themes was performed manually by the author, a qualified pharmacist.
Examples of the text of two DRPs identified by a pharmacist and by MRM in the same patient are shown in Table 1. These DRPs were assigned the theme Hyperlipidemia under/untreated, which captured the basic problem identified within the text of each DRP. MRM Patient has elevated triglycerides and is only taking a statin. Additional treatment, such as a fibrate, may be worth considering Pharmacist Patient’s cholesterol and triglycerides remain elevated despite Lipitor [statin].
This may be due to poor compliance or an inadequate dose
These themes provided a common language for comparison of the DRPs found by the original pharmacist reviewer and MRM. The initial themes were created where at least two of three published prescribing guidelines for the elderly [5, 17, 18] were in agreement concerning the same types of DRPs. DRPs from MRM and pharmacists were mapped to this table of themes. Further themes were added if both pharmacist and MRM DRPs could be mapped to any remaining ‘non-agreement’ prescribing guideline DRPs. New themes were developed for remaining pharmacist and MRM DRPs where concepts were clearly similar but were not contained within prescribing guidelines. These new themes were very broad such as Vitamin, no indication, and Weight Max(F-m) Min(F-m)
Variance binary frequency logF expF frequency achieves the best performance with a F-measure of 0.9647, the highest F-measure of the worst performing schema is 0.9622 (expF), just 0.003% lower than frequency. Furthermore, all weighting schema exhibit the same e ectiveness when considering the worst performing settings. Thus the range of performance di erences and their variance do not signi cantly di er across weighting schema. This may be due to the fact that death certi cates are in general short documents, where features occur uniformly.
Feature stemBigram concept + bigramStem concFullMorph + stemBigram concBigram + stemBigram
concBigram concFullBigram
conceptFull concept + stemBigram concept stem
Feature is the nal variable of our analysis, and the one with the greatest impact on classi cation results. The use of the concFullMorph + stemBigram feature provide the highest F-measure (0.9647), while concFullBigram yields the lowest maximal F-measure (0.7768): a signi cant di erence of 19.48%. The smallest variance was demonstrated by stemBigram (2:02 10 4), making it the most robust feature in our experiment; in addition this feature yielded a maximal Fmeasure of only 0.003% lower than the best value recorded in our experiments. The minimal F-measure yield by the stemBigram feature was also greater than the greatest F-measure values obtained when using half of the features investigated in our study. These results provide strong indication that, of the variables analysed, the choice of feature provides the greatest contribution to the classi cation e ectiveness. 5 Conclusions Timely processing of cancer noti cations is critical for timely reporting of cancer incidence and mortality. Death certi cates are a rich source of data on cancer mortality. Cancer registries acquire free-text death certi cates on a regular (e.g. fortnightly) basis. However, the cause of death information needs to be classi ed to facilitate reporting of cancer mortality. Cause of death information classi ed using ICD-10 codes is only available on an annual basis. In this paper we investigated the automatic classi cation of death certi cates to individuate cancer noti able cause of deaths. The investigated approaches achieved overall strong classi cation e ectiveness, with a support vector machine classi er trained with token bigram features and information from the SNOMED CT medical ontology, and weighted by their frequency in the documents yielding an F-measure of 0.9647. The choice of features, rather than that of classi ers or weighting schema, was found to be the determining factor for high e ectiveness.
Future e orts will be directed towards an in depth error analysis, in particular examining the distance between the prediction produced by a classi er and the decision threshold. We also plan to extend the investigation to predict the actual ICD-10 codes associated to cause of death related to cancer, so as to further assist clinical coders in processing cancer noti cations. Clinician-Driven Automated Classification of Limb Fractures from Free-Text Radiology Reports
Amol Wagholikar1, Guido Zuccon1, Anthony Nguyen1, Kevin Chu2, Shane Martin2, Kim Lai2, Jaimi Greenslade2 Abstract. The aim of this research is to report initial experimental results and evaluation of a clinician-driven automated method that can address the issue of misdiagnosis from unstructured radiology reports. Timely diagnosis and reporting of patient symptoms in hospital emergency departments (ED) is a critical component of health services delivery. However, due to disperse information resources and vast amounts of manual processing of unstructured information, a point-of-care accurate diagnosis is often difficult. A rule-based method that considers the occurrence of clinician specified keywords related to radiological findings was developed to identify limb abnormalities, such as fractures. A dataset containing 99 narrative reports of radiological findings was sourced from a tertiary hospital. The rule-based method achieved an F-measure of 0.80 and an accuracy of 0.80. While our method achieves promising performance, a number of avenues for improvement were identified using advanced natural language processing (NLP) techniques.
Keywords: limb fractures, emergency department, radiology reports, classification, rule-based method, machine learning. 1
Introduction The analysis of x-rays is an essential step in the diagnostic work-up of many conditions including fractures in injured Emergency Department (ED) patients. X-rays are initially interpreted by the treating ED doctor, and if necessary patients are appropriately treated. X-rays are eventually reported on by the specialist in radiology and these findings are relayed to the treating doctor in a formal written report. The ED, however, may not receive the report until after the patient was discharged home. This is not an uncommon event because the reporting did not occur in real-time. As a result, there are potential delays in the diagnosis of subtle fractures missed by the treating doctor until the receipt of the radiologist’s report. The review of x-ray reports is a necessary practice to ensure fractures and other conditions identified by the radiologist were not missed by the treating doctor. The review requires the reading of the free-text report. Large “batches” of x-rays are reviewed often days after the patient’s ED presentation. This is a labour intensive process which adds to the diagnostic delay. The process may be streamlined if it can be automated with clinical text processing solutions. These solutions will minimise delays in diagnosis and prevent complications arising from diagnostic errors [1-2]. This research aims to address these issues through the application of a gazetteer rule-based approach where keywords that may suggest the presence or absence of an abnormality were provided by expert ED clinicians. Rule-based methods are commonly used in Artificial Intelligence [3-5]. Studies have shown that rule-based methods can be applied for identifying clinical conditions from radiology reports such as acute cholecystitis, acute pulmonary embolism and other conditions [6]. The purpose of these methods is to simulate human reasoning for any given information processing task to achieve full or partial automation. 2
Related Work
Previous studies that focused on the problem of identification of subtle limb fractures during the diagnosis of ED patients showed that about 2.1% of all fractures were not identified during initial presentation to the Emergency Department [7]. A similar study about radiological evidence for fracture reports that 1.5% of all x-rays had abnormalities that were not identified in the Emergency Department records [8]. Further research also reported that 5% and 2% of the x-rays of the hand/fingers and ankle/foot from a pediatric Emergency Department had fractures missed by the treating ED doctor [9]. These small percentages of incidences may have significant impact on the overall patient healthcare as these missed fractures may develop into more complex conditions. Timely recognition of fractures is therefore important. There have been efforts to automatically detect fractures and other abnormalities from free-text radiology reports using support vector machine (SVM) and machine learning techniques[10-11]. Even though the results of machine learning based classifiers show high effectiveness, their applicability in clinical settings may be limited. Machine learning methods are data–driven, and as a result, if the training sample is not a representative selection of the problem domain, then the resulting model will not generalise. In addition, machine learning approaches are required to be retrained on new corpora and tasks and collating training data to build new classifier models can be a timely and labour intensive process. These issues provide the motivation for the investigation of rule-based methods which have the ability to model expert knowledge as easily implementable rules. 3
Methods A set of 99 de-identified free-text descriptions of patient’s limb x-rays reported by radiologists were extracted from a tertiary hospital’s picture archiving and communication system (PACS). An ethics approval was granted by the Human Research Ethics Committee at Queensland Health to use this data. The average length of free-text reports is about 52 words with total 930 unique words in the vocabulary. Some reports are semi-structured, with section headings such as “History”, “Clinical Details”, “Findings”, appearing in the text.
Ground Truth Development One ED visiting medical officer and one ED Registrar were engaged as assessors to manually classify the patient findings. Findings were assigned to either one of the following two classes: (1) “Normal”, means identifying no fractures or dislocations and (2) “Abnormal”, identifying the presence of a reportable abnormality such as fracture, dislocation, displacement etc., which requires further follow-up. To gather ground truth labels about the data, an in-house annotation tool was developed. This tool allowed the assessors to manually annotate and classify the free-text reports into one of the two target categories. The two assessors initially agreed on the annotations of 77 of the 99 reports and disagreed on the remaining 22 reports. The disagreed reports were resolved and validated by a senior Staff Specialist in Emergency Medicine, who acted as a third assessor.
Rule-based classifier
A rule-based classifier was developed and implemented with rules as a set of keywords extracted from the x-ray reports assessment criteria as documented by the clinicians prior to the ground truth annotation task. The classifier was implemented to classify the text into “Normal” and “Abnormal” categories as shown in Table 1. Table 1. Keywords used for building the rule-base.
Results and Discussion Results obtained by our gazetteer rule-based approach on the dataset containing 99 radiology reports are reported in Table 2, along with the performance of a Naïve Bayes classifier that was used to classify on the same dataset [12]. The Naïve Bayes classifier was trained and evaluated using a 10-fold cross validation approach. This approach used 90% of reports for training and subsequently evaluated on the remaining 10% within each cross validation fold. The average of the evaluation results across the 10 folds was reported as the classifier’s performance. A set of stemmed tokens in combination with high order semantic features such as SNOMED CT concepts related to morphological abnormalities and disorders generated by the Medtex system [13] were used to represent the reports. Classification results were evaluated in terms of F-measure and accuracy (see Table 2). The number of true positive (TP), true negative (TN), false positive (FP), and false negative (FN) instances were also reported. The rule-based system classified 49 reports as “Normal”. Thirty-three of these were classified as “normal” due to the “no + fracture” rule. The remaining 16 reports did not match any rule, and thus were classified as “normal” (i.e. “no rule fired”). The high false negative count from the rule-based system suggests that the keywords that were used to characterise “Abnormal” cases by the clinician were not complete or adequate to capture all possible cases of abnormalities. Although the proposed keyword rule-based approach is simplistic but shows promise, advanced Natural Language Processing techniques such as those adopted in Medtex [14] can be used to improve classification performances. More keywords can also be learnt using computational linguistic methods, such as the Basilisk bootstrapping algorithm [15]. 5
Conclusion and Future Research This work has described an initial investigation of a clinician-driven rule-based method for automatic classification of free-text limb fracture x-ray findings. We described a simple keyword spotting approach where keywords were derived from classification criteria provided by clinicians. The rule-based classification method achieved promising results with F-measure performances of 0.80 and an accuracy of 0.80. As future work, the research will aim to improve the simple keyword approach with more advanced clinical text processing techniques to complement the proposed rule-based classification method. The possible integration of our method in real-life workflow of hospital emergency departments will also be considered.
Acknowledgements. The authors are thankful to Bevan Koopman for feedbacks on earlier draft of this paper. This research was supported by the Queensland Emergency Medicine Research Foundation Grant, EMPJ-11-158-Chu-Radiology.
Using Prediction to Improve Elective Surgery Scheduling
Zahra Shahabi Kargar1, 2, Sankalp Khanna1, 2, Abdul Sattar1 1 Institute for Integrated and Intelligent Systems, Griffith University, Australia {Zahra.Shahabikargar, A.Sattar@griffith.edu.au 2 The Australian e-Health Research Centre, RBWH, Herston, Australia
{Sankalp.Khanna}@csiro.au Abstract. Stochastic activity durations, uncertainty in the arrival process of patients, and coordination of multiple activities are some key features of surgery planning and scheduling. In this paper we provide an overview of challenges around elective surgery scheduling and propose a predictive model for elective surgery scheduling to be evaluated in a major tertiary hospital in Queensland. The proposed model employs waiting lists, peri-operative information, workload predictions, and improved procedure time estimation models, to optimise surgery scheduling. It is expected that the resulting improvement in scheduling processes will lead to more efficient use of surgical suites, higher productivity, and lower labour costs, and ultimately improve patient outcomes.
Keywords: Surgery scheduling, Predictive optimisation, Waiting list 1 Ageing population and higher rates of chronic disease increase the demand on health services. The Australian Institute of Health and Welfare reports a 3.6% per year increase in total elective surgery admissions over the past four years [1]. These factors stress the need for efficiency and necessitate the development of adequate planning and scheduling systems in hospitals. Since operating rooms (ORs) are the hospital’s largest cost and revenue centre that has a major impact on the performance of the hospital, OR scheduling has been studied by many researchers.
The surgery scheduling problem deals with the allocation of ORs under uncertain demand in a complex and dynamic hospital environment to optimise use of resources. Different techniques such as Mathematical programming[24], simulation [5, 6], Meta-heuristics [5, 7] and Distributed Constraint Optimization [8] have been proposed to address this problem. However most current efforts to solve this problem either make simplifying assumptions (e.g. considering only one department or type of surgery [4]), or employ theoretic data [3, 5] which make them difficult to use in hospitals.
In this paper, we propose a prediction based methodology for surgery scheduling to address the above limitations. By using predicted workload information and retrospective analysis of waiting lists and theatre utilization, we predict a theatre template representing optimal case mix. The proposed model also employs accurate estimation of procedure time and predicted workload information to drive optimal elective surgery scheduling, and help hospitals fulfil National Elective Surgery Targets (NEST) [1]. 2
Elective Surgery Scheduling at the Evaluation Hospital Long waiting lists for elective surgery in Australian hospitals during recent years has driven a nationwide research agenda to improve the planning, management and delivery of health care services. This work is to be evaluated at a major tertiary hospital which has a total of 15 operating theatres performing 124 elective operating sessions and 23 emergency sessions per week. Currently allocation of available elective operating sessions at the hospital have been broken down to different specialties and teams of surgeons based on a static case mix planning. This static allocation of available sessions between emergency and elective patients and among different departments results in underutilization or cancellation due to demand fluctuations. Also, the allocation of patients to theatres is carried out without considering the uncertainty and possible changes that might happen. Procedure times are estimated by using generic data or recommended by relevant surgeons not based on individual patient and surgery characteristics. Patients are booked into schedules in a joint process between surgeons and the booking department. Due to the dynamic environment and rapid changes, these schedules need to be updated quickly. Usually department managers have regular meetings to make any changes needed. Department managers try to locally optimise their department goals, but since there is no global objective usually these solutions are not the optimal global solutions. 3
An Optimal Surgery Scheduling Model Although the surgery scheduling problem has been well addressed in literature, it still remains an open problem in Operations Research and Artificial Intelligence. Despite the dynamic nature of the hospital environment, the majority of previous studies ignore the underlying uncertainty. This leads to simplistic models that are not applicable in real world situations.
Current State of the Art Cardoen et al. present a comprehensive literature review on operating room scheduling including different features such as performance measures, patient classes, solution technique and uncertainty [9]. One of the major issues associated with the development of accurate operating room schedules or capacity planning strategies is the uncertainty inherent to surgical services. Uncertainty and variability of frequency and distribution of patient arrivals, patient conditions, and procedure durations, as well as ‘‘add-on’’ cases are some instances of uncertainty in surgery scheduling [10]. Among them stochastic arrival and procedure duration are two type of uncertainty studied by many researchers. Procedure duration depends on several factors such as experience of the surgeon, supporting staff, type of anaesthesia, and precondition of the patient. Devi et al. estimate surgery times by using Adaptive Nero Fuzzy Inference Systems, Artificial Neural Networks and Multiple Linear Regression Analysis [2] but they just focus on one department and use a very limited sample to build and validate their model. Lamiri et al. developed a stochastic model for planning elective surgeries under uncertain demand for emergency surgery [3]. Lamiri et al. also address the elective surgery planning under uncertainties related to surgery times and emergency surgery demands by combining Monte Carlo simulation and a column generation approach[5]. Although their method addresses uncertainties, it is based on theoretic data and it has not been tested on real data. What is needed is a whole of theatre approach to provide better prediction of surgery time, incorporation of predicted workload in planning the weekly surgery template, and target guided optimization to ensure optimal allocation of resources. 3.2 To improve the planning and optimization tasks underlying the process, we propose a two stage methodology for elective surgery scheduling. As a first stage, predicted workload information (drawn from Patient Admission Prediction Tool [11] currently used at the evaluation hospital), current Waiting List information and Historic utilization information is used to manage theatre allocation and case mix distribution for each week (see Figure 1). This allows the prediction based sharing of theatres between elective and emergency surgery, and allocation of theatre time to surgery teams/departments and results in a theatre schedule template that works better than a static allocation model (as demonstrated by Khanna et al. [8]). In the second stage of the process, the allocation of patients to the weekly theatre schedule is guided by an improved prediction algorithm to estimate the surgery duration. The algorithm takes into account current patient, surgery, and surgeon information and related historic peri-operative information to forecast the planned procedure time. Incorporating NEST compliance in the optimization function and improved resource estimation deliver further improvements to the scheduling process and help deliver a more robust and optimal schedule (Figure 1). We are currently working towards collecting over 5 years of surgery scheduling, waiting list and peri-operative information for the evaluation hospital from the corporate information systems. This data will be used for modelling and independently validating the prediction algorithms and building historic resource utilization knowledge banks to guide other stage of the scheduling process. 4
Conclusion The proposed model has the potential to improve elective surgery scheduling by providing more accurate procedure time estimation and predicting arrival demand of elective and emergency patients.
If you fire together, you wire together
Prajni Sadananda1, Ramakoti Sadananda2, 3 1 Department of Anatomy and Neuroscience, University of Melbourne Australian, Australia prajni.sadananda@unimelb.edu.au 2 Institute for Integrated and Intelligent Systems, Griffith University, Australia 3NICTA, Sydney Australia rsadananda@griffith.edu.au
The intention of this paper is to stimulate discussion on Hebb’s Law and its pedagogic implications.
At a basic cellular level, Hebb’s Law states that is Cell A and Cell B persis-‐ tently fire, the connection between them strengthens. Figure 1 illustrates the interactions. This is a cellular levels process, suggesting that brain pro-‐ cesses that occur repeartedly tend to become grafted together [1]. Fig 1. Hebb’s Law. Repeated stimulation results in a stronger signal
This scientific theory explains the adaptation of neurons during the learn-‐ ing process. Importantly, this type of plasticity does not involve increasing the number of cells, but rather strengthening the existing cells’ connectivity. Understanding such biological phenomena opens up new paradigms and laws that AI can utilise. The question is whether Hebb’s law would stand at a higher level of abstraction? There are suggestive, but not conclusive indica-‐ tions. For example, a friendship is considered stronger with time, indicating a strengthening of wiring between the friends.
Models based on “firing together to wire together” have been suggested in health and therapy [2]. For example, if a patient presents with a mental trauma that causes extreme anger, the therapist introduces a counter and positive stimulus that occurs whenever the anger occurs. Both (anger and the positive stimulus) are repeated over and over again, thus following Hebb’s Law and adding strength to this connection between the two stimuli, resulting in relief to the patient.
This also implies causal and temporal conjectures based on causality. The causality is in the firing sequence; that if A fires first and then B fires, A is the cause. If B fires before A, a reverse interpretation is possible that may de-‐ crease the strength between them. There seems evidence to suggest that the “firing” and “wiring” may be a sequential process.
Causality is a subject of intense philosophical interest from ancient times. Most causal models are rule-‐based systems. They demand descriptions of the world at two points in time – a before and an after. Two problems arise here: the practical computational compulsions make these rules crudely simplistic. In addition, it is challenging to incorporate temporal effects within the framework of rule based systems. Hebb’s law, while suggesting causali-‐ ty, does not provide any quantification. Thus, it is unlikely that an alternative formulation of causation would emerge from Hebb’s law alone. We may look for another, additional neural network perspective of causation here.
Nevertheless, causality as implied with Hebb’s law has been used in sci-‐ entific research and therapeutics to a large’ extent. For example, oftentimes doctors complain of their patients being unable to add minor and incremen-‐ tal changes in their daily routines (such as exercise). Understanding Hebb’s law will open new insights into why this might be so. It is possible that the patient is not yet “wired” in this activity and requires more “firing” before these changes can be established. An avenue for AI research is to aide in the development of tools to help such people to “re-‐wire”.
Indeed, such tools exist to some extent to treat spinal cord injured pa-‐ tients who have lost motor control of their limbs. In a non-‐injured situation, the brain delivers pulses to the lower limbs in a rhythmic/patterned fashion to allow walking action. Once a spinal injury occurs, the connectivity from the brain to the limbs is lost, thereby leaving the patient immobile. Stimula-‐ tors are often placed below the level of the injury, which deliver patterned pulses in a similar manner to what the brain was previously doing. Over a period of time, a spinal pattern generator emerges, which thus allows some motion of the lower limbs [3]. This area of research is as yet in its infancy and calls for a better, more intelligent systems to aide these patients. Conclusions:
Artificial intelligence in health opens up chapters of great opportunities and exciting challenges. The logical calculus articulated by McCulloch and Pitts [4] forms the initial basis for both Symbolic and Connectionist AI. Since then a number of paradigms have emerged on all aspects of AI and relating to health and health care. The emergence of the convergence of computing and communication provides us boundless opportunities to exploit these paradigms and discover the new ones. ReferenceƐ:
Hebb, D. O.: Organization of Behavior: a Neuropsychological Theory. John Wiley, New York (1949). Atkinson, B., Atkinson, L., Kutz, P., Lata, L., Lata, K.W., Szekely, J., Weiss, P.: Rewiring Neural States in Couples Therapy: Advances from Affective Neuroscience. In: Journal of Systemic Therapies. 24, 3-‐13 (2005) Edgerton, V.R., Roy, R.R.: A new age for rehabilitation. Eur J Phys Re-‐ habil Med. 48, 99-‐109 (2012) McCulloch, W.S., Pitts, W.: A logical Calculus of the ideas immanent in nervous activity, Bulletin of mathematical Biophysics. 5, 115-‐137 (1943).