Developing and validating a parser-based suicidality detection model in text-based mental health services.

BACKGROUND: Advances in text-mining can potentially aid online text-based mental health services in detecting suicidality. However, false positive remains a challenge. METHODS: Data of a free 24/7 online text-based counseling service in Hong Kong were used to develop a novel parser-based algorithm (PBSD) to detect suicidal ideation while minimizing false alarms. Sessions containing keywords related to suicidality were extracted (N = 1267). PBSD first applies a sentence parser to work out the grammatical structure of each sentence, including subject, object, dependent and modifier. Then a set of syntax rules were applied to judge if a flagged sentence is a true or false positive. Half of the sessions were randomly selected to train PBSD, the remaining were used as the test set. A standard keywords matching model was adopted as the baseline comparison. Accuracy and recall were reported to demonstrate models' performance. RESULTS: Of the 1267 sessions, 585 (46.2 %) were false alarms. The false alarms were categorized into four types: negation-induced false alarms (NIFA; 14 %), subject-induced false alarms (SIFA; 19 %), tense-induced false alarms (TIFA; 30 %), and other types of false alarms (OTFA; 37 %). PBSD significantly outperforms the baseline keywords matching model on accuracy (0.68 vs 0.53, 28.3 %). It successfully amended 36.8 % (105/297) lexicon matching-caused false alarms. The reduction on recall was marginal (1 vs 0.96, 4 %). CONCLUSIONS: The proposed model significantly improves the use of lexicon-based method by reducing false alarms and improving the accuracy of suicidality detection. It can potentially reduce unnecessary panic and distraction caused by false alarms among frontline service-providers.

Social, lifestyle, and health status characteristics as a proxy for occupational burnout identification: A network approach analysis.

Background: Occupational burnout is a type of psychological syndrome. It can lead to serious mental and physical disorders if not treated in time. However, individuals tend to conceal their genuine feelings of occupational burnout because such disclosures may elicit bias from superiors. This study aims to explore a novel method for estimating occupational burnout by elucidating its links with social, lifestyle, and health status factors. Methods: In this study 5,794 participants were included. Associations between occupational burnout and a set of features from a survey was analyzed using Chi-squared test and Wilcoxon rank sum test. Variables that are significantly related to occupational burnout were grouped into four categories: demographic, work-related, health status, and lifestyle. Then, from a network science perspective, we inferred the colleague's social network of all participants based on these variables. In this inferred social network, an exponential random graph model (ERGM) was used to analyze how occupational burnout may affect the edge in the network. Results: For demographic variables, age (p < 0.01) and educational background (p < 0.01) were significantly associated with occupational burnout. For work-related variables, type of position (p < 0.01) was a significant factor as well. For health and chronic diseases variables, self-rated health status, hospitalization history in the last 3 years, arthritis, cardiovascular diseases, high blood lipid, breast diseases, and other chronic diseases were all associated with occupational burnout significantly (p < 0.01). Breakfast frequency, dairy consumption, salt-limiting tool usage, oil-limiting tool usage, vegetable consumption, pedometer (step counter) usage, consuming various types of food (in the previous year), fresh fruit and vegetable consumption (in the previous year), physical exercise participation (in the previous year), limit salt consumption, limit oil consumption, and maintain weight were also significant factors (p < 0.01). Based on the inferred social network among all airport workers, ERGM showed that if two employees were both in the same occupational burnout status, they were more likely to share an edge (p < 0.0001). Limitation: The major limitation of this work is that the social network for occupational burnout ERGM analysis was inferred based on associated factors, such as demographics, work-related conditions, health and chronic diseases, and behaviors. Though these factors have been proven to be associated with occupational burnout, the results inferred by this social network cannot be warranted for accuracy. Conclusion: This work demonstrated the feasibility of identifying people at risk of occupational burnout through an inferred colleague's social network. Encouraging staff with lower occupational burnout status to communicate with others may reduce the risk of burnout for other staff in the network.

Predicting Acute Onset of Heart Failure Complicating Acute Coronary Syndrome: An Explainable Machine Learning Approach.

Patients with acute coronary syndrome (ACS) are at high risk of heart failure (HF). Early prediction and management of HF among ACS patients are essential to provide timely and cost-effective care. The aim of this study is to train and evaluate a machine learning model to predict the acute onset of HF subsequent to ACS. A total of 1,028 patients with ACS admitted to Guangdong Second Provincial General Hospital between October 2019 and May 2022 were included in this study. 128 clinical features were ranked using Shapley additive exPlanations (SHAP) values and the top 20% of features were selected for building a balanced random forest (BRF) model. We compared the discriminatory capability of BRF with linear logistic regression (LLR). In the hold-out test set, the BRF model predicted subsequent HF with areas under the curve (AUC) of 0.76 (95% CI: 0.75-0.77), sensitivity of 0.97 (95% CI: 0.96-0.97), positive predictive value (PPV) of 0.73 (95% CI: 0.72-0.74), negative predictive value (NPV) of 0.63 (95% CI: 0.60-0.66), and accuracy of 0.73 (95% CI: 0.72-0.73), respectively. BRF outperforms linear logistic regression by 15.6% in AUC, 3.0% in sensitivity, and 60.8% in NPV. End-to-end machine learning approaches can predict the acute onset of HF following ACS with high prediction accuracy. This proof-of-concept study has the potential to substantially advance the management of ACS patients by utilizing the machine learning model as a triage tool to automatically identify clinically significant patients allowing for prioritization of interventions.

Fatal crashes and rare events logistic regression: an exploratory empirical study.

Objective: Fatal road accidents are statistically rare, posing challenges for accurate estimation through the classic logit model (LM). This study seeks to validate the efficacy of a rare events logistic model (RELM) in enhancing the precision of fatal crash estimations. Methods: Both LM and RELM were employed to examine the relationship between pertinent risk factors and the incidence of fatal crashes. Crash-injury datasets sourced from Hillsborough County, Florida served as the empirical basis for evaluating the performance metrics of both LM and RELM. Results: The analysis revealed that RELM yielded more accurate predictions of fatal crashes compared to LM. Receiver operating characteristic (ROC) curves were constructed, and the area under the curve (AUC) for each model was computed to offer a comparative performance assessment. The empirical evidence notably favored RELM over LM as substantiated by superior AUC values. Conclusion: The study offers empirical validation that RELM is demonstrably more proficient in predicting fatal crashes than the LM, thereby recommending its application for nuanced traffic safety analytics.

Network-based prediction of the disclosure of ideation about self-harm and suicide in online counseling sessions.

BACKGROUND: In psychological services, the transition to the disclosure of ideation about self-harm and suicide (ISS) is a critical point warranting attention. This study developed and tested a succinct descriptor to predict such transitions in an online synchronous text-based counseling service. METHOD: We analyzed two years' worth of counseling sessions (N = 49,770) from Open Up, a 24/7 service in Hong Kong. Sessions from Year 1 (N = 20,618) were used to construct a word affinity network (WAN), which depicts the semantic relationships between words. Sessions from Year 2 (N = 29,152), including 1168 with explicit ISS, were used to train and test the downstream ISS prediction model. We divided and classified these sessions into ISS blocks (ISSBs), blocks prior to ISSBs (PISSBs), and non-ISS blocks (NISSBs). To detect PISSB, we adopted complex network approaches to examine the distance among different types of blocks in WAN. RESULTS: Our analyses find that words within a block tend to form a module in WAN and that network-based distance between modules is a reliable indicator of PISSB. The proposed model yields a c-statistic of 0.79 in identifying PISSB. CONCLUSIONS: This simple yet robust network-based model could accurately predict the transition point of suicidal ideation prior to its explicit disclosure. It can potentially improve the preparedness and efficiency of help-providers in text-based counseling services for mitigating self-harm and suicide.

In online counseling, the help-provider can often be engaging with several service users simultaneously. Therefore, new tools that could help to alert and assist the help-provider and increase their preparedness for getting further help for service users could be useful. In this study, we developed and tested a new tool that is designed to alert help-providers to the disclosure of self-harm and suicidal thoughts, based on the words that the service user has been typing. The tool is developed on the basis that word usage may have a specific pattern when suicidal thoughts are more likely to occur. We tested our tool using two years' worth of online counseling conversations and we show that our approach can help to predict the confession of suicidal thoughts. As such, we are taking a step forward in helping to improve these counseling services.

Deep learning identifies explainable reasoning paths of mechanism of action for drug repurposing from multilayer biological network.

The discovery and repurposing of drugs require a deep understanding of the mechanism of drug action (MODA). Existing computational methods mainly model MODA with the protein-protein interaction (PPI) network. However, the molecular interactions of drugs in the human body are far beyond PPIs. Additionally, the lack of interpretability of these models hinders their practicability. We propose an interpretable deep learning-based path-reasoning framework (iDPath) for drug discovery and repurposing by capturing MODA on by far the most comprehensive multilayer biological network consisting of the complex high-dimensional molecular interactions between genes, proteins and chemicals. Experiments show that iDPath outperforms state-of-the-art machine learning methods on a general drug repurposing task. Further investigations demonstrate that iDPath can identify explicit critical paths that are consistent with clinical evidence. To demonstrate the practical value of iDPath, we apply it to the identification of potential drugs for treating prostate cancer and hypertension. Results show that iDPath can discover new FDA-approved drugs. This research provides a novel interpretable artificial intelligence perspective on drug discovery.

Age-related changes in the risk of high blood pressure.

Background and aims: Understanding the age-related trend of risk in high blood pressure (BP) is important for preventing heart failure and cardiovascular diseases. But such a trend is still underexplored. This study aims to (a) depict the relationship of BP patterns with age, and (b) understand the trend of high BP prevalence over time in different age groups. Materials and methods: Health check-up data with an observational period of 8 years (January 1, 2011, to December 31, 2018) was used as the data source. A total of 71,468 participants aged over 18 years old with complete information on weight, height, age, gender, glucose, triglyceride, total cholesterol, systolic (SBP), and diastolic blood pressure (DBP) were included for analysis. Generalized additive models were adopted to explore the relationship between the risk of high BP and age. Variance analysis was conducted by testing the trend of high BP prevalence in age groups over time. Results: Risk of high SBP showed a continuous rise from age 35 to 79 years and a concurrent early increase in the risk of high DBP; after age 50-65 years, high DBP risk declined. The risk of SBP rises linearly with age for men, whereas increases non-linearly for women. In addition, a significant increasing trend of high SBP risk among middle-aged people was found during the past decade, men experienced a later but longer period of increase in high SBP than women. Conclusion: The high SBP risk progresses more rapidly in the early lifetime in women, compared to the lifetime thereafter. Thresholds of increasing trend of SBP suggest a possible need for hypertension screening in China after the age of 40.

A Comorbidity Knowledge-Aware Model for Disease Prognostic Prediction.

Prognostic prediction is the task of estimating a patient's risk of disease development based on various predictors. Such prediction is important for healthcare practitioners and patients because it reduces preventable harm and costs. As such, a prognostic prediction model is preferred if: 1) it exhibits encouraging performance and 2) it can generate intelligible rules, which enable experts to understand the logic of the model's decision process. However, current studies usually concentrated on only one of the two features. Toward filling this gap, in the present study, we develop a novel knowledge-aware Bayesian model taking into consideration accuracy and transparency simultaneously. Real-world case studies based on four years' territory-wide electronic health records are conducted to test the model. The results show that the proposed model surpasses state-of-the-art prognostic prediction models in accuracy and c-statistic. In addition, the proposed model can generate explainable rules.

Modelling the geographical spread of HIV among MSM in Guangdong, China: a metapopulation model considering the impact of pre-exposure prophylaxis.

Men who have sex with men (MSM) make up the majority of new human immunodeficiency virus (HIV) diagnoses among young people in China. Understanding HIV transmission dynamics among the MSM population is, therefore, crucial for the control and prevention of HIV infections, especially for some newly reported genotypes of HIV. This study presents a metapopulation model considering the impact of pre-exposure prophylaxis (PrEP) to investigate the geographical spread of a hypothetically new genotype of HIV among MSM in Guangdong, China. We use multiple data sources to construct this model to characterize the behavioural dynamics underlying the spread of HIV within and between 21 prefecture-level cities (i.e. Guangzhou, Shenzhen, Foshan, etc.) in Guangdong province: the online social network via a gay social networking app, the offline human mobility network via the Baidu mobility website, and self-reported sexual behaviours among MSM. Results show that PrEP initiation exponentially delays the occurrence of the virus for the rest of the cities transmitted from the initial outbreak city; hubs on the movement network, such as Guangzhou, Shenzhen, and Foshan are at a higher risk of 'earliest' exposure to the new HIV genotype; most cities acquire the virus directly from the initial outbreak city while others acquire the virus from cities that are not initial outbreak locations and have relatively high betweenness centralities, such as Guangzhou, Shenzhen and Shantou. This study provides insights in predicting the geographical spread of a new genotype of HIV among an MSM population from different regions and assessing the importance of prefecture-level cities in the control and prevention of HIV in Guangdong province. This article is part of the theme issue 'Data science approach to infectious disease surveillance'.

Development of an interpretable machine learning-based intelligent system of exercise prescription for cardio-oncology preventive care: A study protocol.

Background: Cardiovascular disease (CVD) and cancer are the first and second causes of death in over 130 countries across the world. They are also among the top three causes in almost 180 countries worldwide. Cardiovascular complications are often noticed in cancer patients, with nearly 20% exhibiting cardiovascular comorbidities. Physical exercise may be helpful for cancer survivors and people living with cancer (PLWC), as it prevents relapses, CVD, and cardiotoxicity. Therefore, it is beneficial to recommend exercise as part of cardio-oncology preventive care. Objective: With the progress of deep learning algorithms and the improvement of big data processing techniques, artificial intelligence (AI) has gradually become popular in the fields of medicine and healthcare. In the context of the shortage of medical resources in China, it is of great significance to adopt AI and machine learning methods for prescription recommendations. This study aims to develop an interpretable machine learning-based intelligent system of exercise prescription for cardio-oncology preventive care, and this paper presents the study protocol. Methods: This will be a retrospective machine learning modeling cohort study with interventional methods (i.e., exercise prescription). We will recruit PLWC participants at baseline (from 1 January 2025 to 31 December 2026) and follow up over several years (from 1 January 2027 to 31 December 2028). Specifically, participants will be eligible if they are (1) PLWC in Stage I or cancer survivors from Stage I; (2) aged between 18 and 55 years; (3) interested in physical exercise for rehabilitation; (4) willing to wear smart sensors/watches; (5) assessed by doctors as suitable for exercise interventions. At baseline, clinical exercise physiologist certificated by the joint training program (from 1 January 2023 to 31 December 2024) of American College of Sports Medicine and Chinese Association of Sports Medicine will recommend exercise prescription to each participant. During the follow-up, effective exercise prescription will be determined by assessing the CVD status of the participants. Expected outcomes: This study aims to develop not only an interpretable machine learning model to recommend exercise prescription but also an intelligent system of exercise prescription for precision cardio-oncology preventive care. Ethics: This study is approved by Human Experimental Ethics Inspection of Guangzhou Sport University. Clinical trial registration: http://www.chictr.org.cn, identifier ChiCTR2300077887.

The Opportunities and Challenges of the First Three Years of Open Up, an Online Text-Based Counselling Service for Youth and Young Adults.

We present the opportunities and challenges of Open Up, a free, 24/7 online text-based counselling service to support youth in Hong Kong. The number of youths served more than doubled within the first three years since its inception in 2018 in response to increasing youth suicidality and mental health needs. Good practice models are being developed in order to sustain and further scale up the service. We discuss the structure of the operation, usage pattern and its effectiveness, the use of AI to improve users experience, and the role of volunteer in the operation. We also present the challenges in further enhancing the operation, calling for more research, especially on the identification of the optimal number of users that can be concurrently served by a counsellor, the effective approach to respond to a small percentage of repeated users who has taken up a disproportional volume of service, and the way to optimize the use of big data analytics and AI technology to enhance the service. These advancements will benefit not only Open Up but also similar services across the globe.

Detecting premature departure in online text-based counseling using logic-based pattern matching.

BACKGROUND: More so than face-to-face counseling, users of online text-based services might drop out from a session before establishing a clear closure or expressing the intention to leave. Such premature departure may be indicative of heightened risk or dissatisfaction with the service or counselor. However, there is no systematic way to identify this understudied phenomenon. PURPOSE: This study has two objectives. First, we developed a set of rules and used logic-based pattern matching techniques to systematically identify premature departures in an online text-based counseling service. Second, we validated the importance of premature departure by examining its association with user satisfaction. We hypothesized that the users who rated the session as less helpful were more likely to have departed prematurely. METHOD: We developed and tested a classification model using a sample of 575 human-annotated sessions from an online text-based counseling platform. We used 80% of the dataset to train and develop the model and 20% of the dataset to evaluate the model performance. We further applied the model to the full dataset (34,821 sessions). We compared user satisfaction between premature departure and completed sessions based on data from a post-session survey. RESULTS: The resulting model achieved 97% and 92% F1 score in detecting premature departure cases in the training and test sets, respectively, suggesting it is highly consistent with the judgment of human coders. When applied to the full dataset, the model classified 15,150 (43.5%) sessions as premature departure and the remaining 19,671 (56.5%) as completed sessions. Completed cases (15.2%) were more likely to fill the post-chat survey than premature departure cases (4.0%). Premature departure was significantly associated with lower perceived helpfulness and effectiveness in distress reduction. CONCLUSIONS: The model is the first that systematically and accurately identifies premature departure in online text-based counseling. It can be readily modified and transferred to other contexts for the purpose of risk mitigation and service evaluation and improvement.

GraphSynergy: a network-inspired deep learning model for anticancer drug combination prediction.

OBJECTIVE: To develop an end-to-end deep learning framework based on a protein-protein interaction (PPI) network to make synergistic anticancer drug combination predictions. MATERIALS AND METHODS: We propose a deep learning framework named Graph Convolutional Network for Drug Synergy (GraphSynergy). GraphSynergy adapts a spatial-based Graph Convolutional Network component to encode the high-order topological relationships in the PPI network of protein modules targeted by a pair of drugs, as well as the protein modules associated with a specific cancer cell line. The pharmacological effects of drug combinations are explicitly evaluated by their therapy and toxicity scores. An attention component is also introduced in GraphSynergy, which aims to capture the pivotal proteins that play a part in both PPI network and biomolecular interactions between drug combinations and cancer cell lines. RESULTS: GraphSynergy outperforms the classic and state-of-the-art models in predicting synergistic drug combinations on the 2 latest drug combination datasets. Specifically, GraphSynergy achieves accuracy values of 0.7553 (11.94% improvement compared to DeepSynergy, the latest published drug combination prediction algorithm) and 0.7557 (10.95% improvement compared to DeepSynergy) on DrugCombDB and Oncology-Screen datasets, respectively. Furthermore, the proteins allocated with high contribution weights during the training of GraphSynergy are proved to play a role in view of molecular functions and biological processes, such as transcription and transcription regulation. CONCLUSION: The introduction of topological relations between drug combination and cell line within the PPI network can significantly improve the capability of synergistic drug combination identification.

Detecting suicide risk using knowledge-aware natural language processing and counseling service data.

RATIONALE: Detecting users at risk of suicide in text-based counseling services is essential to ensure that at-risk individuals are flagged and prioritized. OBJECTIVE: The objective of this study is to develop a domain knowledge-aware risk assessment (KARA) model to improve our ability of suicide detection in online counseling systems. METHODS: We obtained the largest known de-identified dataset from an emotional support system established in Hong Kong, comprising 5682 Cantonese conversations between help-seekers and counselors. Of those, 682 conversations disclosed crisis intentions of suicide. We constructed a suicide-knowledge graph, representing suicide-related domain knowledge as a computer-processible graph. Such knowledge graph was embedded into a deep learning model to improve its ability to identify help-seekers in crisis. As the baseline, a standard NLP model was applied to the same task. 80% of the study samples were randomly sampled to train model parameters. The remaining 20% were used for model validation. Evaluation metrics including precision, recall, and c-statistic were reported. RESULTS: Both KARA and the baseline achieved high precision (0.984 and 0.951, shown in Table 2) and high recall (0.942 and 0.947) towards non-crisis cases. For crisis cases, however, KARA model achieved a much higher recall than the baseline (0.870 vs 0.791). The c-statistics of KARA and the baseline were 0.815 and 0.760, respectively. CONCLUSION: KARA significantly outperformed standard NLP models, demonstrating good translational value and clinical relevance.

Risk of suicide after a self-poisoning episode: a self-controlled case series study.

PURPOSE: The risk of death from suicide after self-poisoning has been known to be significantly higher as compared to the general population. Nevertheless, the change in suicide risk before and after self-poisoning is still unclear. METHODS: The database of territory-wide medical records collected by the Hospital Authority of Hong Kong was used to identify inpatients who have survived the first-ever self-poisoning but died by suicide between January 1, 2001, and December 31, 2010. A self-controlled case series ("SCCS") design controlling for time-invariant patient confounders was used to explore the temporal change in suicide risk after the first self-poisoning episode. RESULTS: During the study period, 227 people in the database died from suicide after surviving one episode of self-poisoning. A significant increase of the risk of suicide in the first 12 months after the first lifetime self-poisoning-Risk Ratio ("RR") 2.88 (95% CI 1.74-4.76)-was detected. The RR gradually returned to baseline levels after the second post-poisoning period. CONCLUSION: By within-person comparison, the net increase of the suicide risk caused by the first self-poisoning was quantitatively modeled, demonstrating that the first lifetime self-poisoning event itself is a modifiable risk factor of subsequent suicide death.

Understanding the Association Between Antidepressants and the Risk of Being Diagnosed with Dementia in Older People: A Self-Controlled Case Series Study.

BACKGROUND: Given concerns about adverse outcomes for older people taking antidepressants in the literature, we investigated whether taking antidepressants elevates the risk of dementia. OBJECTIVE: This study aims to investigate the putative association of antidepressants with the risk of dementia. METHODS: We conducted a population-based self-controlled case series analysis of older people with dementia and taking antidepressants, using territory-wide medical records of 194,507 older patients collected by the Hospital Authority of Hong Kong, to investigate the association between antidepressant treatment and the risk of developing dementia in older people. RESULTS: There was a significantly higher risk of being diagnosed with dementia during the pre-drug-exposed period (incidence rate ratio (IRR) 20.42 (95% CI: 18.66-22.34)) compared to the non-drug-exposed baseline period. The IRR remained high during the drug-exposed period (IRR 8.86 (7.80-10.06)) before returning to a baseline level after washout (IRR 1.12 (0.77-1.36)). CONCLUSION: The higher risk of dementia before antidepressant treatment may be related to emerging psychiatric symptoms co-occurring with dementia, which trigger medical consultations that result in a decision to begin antidepressants. Our findings do not support a causal relationship between antidepressant treatment and the risk of dementia.

Usefulness of preoperative coronary computed tomography angiography in high risk non-cardiovascular surgery old patients with unknown or suspected coronary artery disease.

BACKGROUND: Cumulative evidence has shown that the non-invasive modality of coronary computed tomography angiography (CCTA) has evolved as an alternative to invasive coronary angiography, which can be used to quantify plaque burden and stenosis and identify vulnerable plaque, assisting in diagnosis, prognosis and treatment. With the increasing elderly population, many patients scheduled for non-cardiovascular surgery may have concomitant coronary artery disease (CAD). The aim of this study was to investigate the usefulness of preoperative CCTA to rule out or detect significant CAD in this cohort of patients and the impact of CCTA results to clinical decision-making. METHODS: 841 older patients (age 69.5 ± 5.8 years, 74.6% males) with high risk non-cardiovascular surgery including 771 patients with unknown CAD and 70 patients with suspected CAD who underwent preoperative CCTA were retrospectively enrolled. Multivariate logistic regression analysis was performed to determine predictors of significant CAD and the event of cancelling scheduled surgery in patients with significant CAD. RESULTS: 677 (80.5%) patients had non-significant CAD and 164 (19.5%) patients had significant CAD. Single-, 2-, and 3- vessel disease was found in 103 (12.2%), 45 (5.4%) and 16 (1.9%) patients, respectively. Multivariate analysis demonstrated that positive ECG analysis and Agatston score were independently associated with significant CAD, and the optimal cutoff of Agatston score was 195.9. The event of cancelling scheduled surgery was increased consistently according to the severity of stenosis and number of obstructive major coronary artery. Multivariate analysis showed that the degree of stenosis was the only independent predictor for cancelling scheduled surgery. In addition, medication using at perioperative period increased consistently according to the severity of stenosis. CONCLUSIONS: In older patients referred for high risk non-cardiovascular surgery, preoperative CCTA was useful to rule out or detect significant CAD and subsequently influence patient disposal. However, it might be unnecessary for patients with negative ECG and low Agatston score. Trial registration Retrospectively registered.

Predicting post-discharge self-harm incidents using disease comorbidity networks: A retrospective machine learning study.

BACKGROUND: Self-harm is preventable if the risk can be identified early. The co-occurrence of multiple diseases is related to self-harm risk. This study develops a comorbidity network-based deep learning framework to improve the prediction of individual self-harm. METHODS: Between 01/01/2007-12/31/2010, we obtained 2,323 patients with self-harm records and 46,460 randomly sampled controls from 1,764,094 inpatients across 44 public hospitals in Hong Kong. 80% of the samples were randomly selected for model training, and the remaining 20% were set aside for model testing. We propose a novel patient embedding method, namely Dx2Vec (Diagnoses to Vector), based on the comorbidity network constructed by all historical diagnoses. Dx2Vec represents the comorbidity patterns among diseases and temporal patterns of historical admissions for each patient. RESULTS: Experiments demonstrate that the Dx2Vec-based model outperforms the baseline deep learning model in identifying patients who would self-harm within 12 months (C-statistic: 0.89). The precision is 0.54 for positive cases and 0.98 for negative cases, whilst the recall is 0.72 for positive cases and 0.96 for negative cases. The model extracted the most predictive diagnoses, and pairwise comorbid diagnoses to help medical professionals identify patients with risk. LIMITATIONS: The inpatient data does not contain lab test information. CONCLUSIONS: Incorporation of a disease comorbidity network can significantly improve self-harm prediction performance, indicating that it is critical to consider comorbidity patterns in self-harm screening and prevention programs. The findings have the potential to be translated into effective self-harm screening systems.